#### 01. An education startup is developing an AI-powered tutor to assist middle and high school students with math problems. The goal is to generate accurate, step-by-step solutions to math questions asked in natural language without retraining the underlying large language model. The team is exploring strategies to improve the model's reasoning and explanation abilities, especially for multi-step algebra and geometry problems, while maintaining system scalability and avoiding expensive model retraining. They are considering various prompting and retrieval techniques to guide the model's responses effectively. Which solution best addresses this requirement? - Use zero-shot prompting with minimal input to reduce inference latency. - Implement retrieval-augmented generation (RAG) to fetch similar solved problems from a database. - Fine-tune the model using historical Q&A pairs from previous tutoring sessions. - Use chain-of-thought prompting to guide the model through logical reasoning steps. **CORRECT:** "Use chain-of-thought prompting to guide the model through logical reasoning steps" is the correct answer. Chain-of-thought prompting is a method that encourages a large language model to reason step-by-step before arriving at a final answer. Instead of generating a direct response, the model is prompted to break down the problem into smaller, logical parts and explain each step in the reasoning process. This is especially useful for multi-step problems in subjects like math, where understanding the process is just as important as the final answer. For the education startup, this approach improves the clarity and accuracy of explanations without needing to retrain the model. It also keeps the system scalable and cost-effective since prompting techniques can be applied at runtime using existing foundation models like those available in Amazon Bedrock. **INCORRECT:** "Use zero-shot prompting with minimal input to reduce inference latency" is incorrect. Zero-shot prompting involves asking a model to solve a problem without giving examples or additional guidance. While it may reduce latency, it typically leads to less accurate responses, especially for complex, multi-step math problems. The lack of reasoning steps makes it unsuitable for an educational tool aiming to teach concepts clearly. **INCORRECT:** "Implement retrieval-augmented generation (RAG) to fetch similar solved problems from a database" is incorrect. RAG involves retrieving relevant documents or examples to assist the model in generating responses. Although helpful for providing context, RAG adds complexity and may not always return closely matching examples, especially for unique or newly phrased problems. It also doesn't guarantee step-by-step reasoning. **INCORRECT:** "Fine-tune the model using historical Q&A pairs from previous tutoring sessions" is incorrect. Fine-tuning improves performance in domain-specific tasks but requires labeled training data, compute resources, and infrastructure for model training and deployment. Since the startup wants to avoid retraining for cost and scalability reasons, this option does not meet their requirements. **References:** https://docs.aws.amazon.com/nova/latest/userguide/prompting-chain-of-thought.html https://docs.aws.amazon.com/bedrock/latest/userguide/what-is-prompt-engineering.html Domain: Applications of Foundation Models --- #### 02. A news agency wants to develop an AI assistant that retrieves relevant news articles and generates concise summaries in real time. They are considering Amazon Bedrock for implementing a fully managed Retrieval Augmented Generation (RAG) workflow to enhance the assistant's accuracy and relevance. Which approach would be most effective for achieving this goal in Amazon Bedrock? - Using Amazon Bedrock Guardrails - Using Amazon S3 with Amazon Bedrock - Using Knowledge Bases for Amazon Bedrock - Fine-tuning Amazon Bedrock with predefined prompts **CORRECT:** "Using Knowledge Bases for Amazon Bedrock" is the correct answer. Amazon Bedrock Knowledge Bases is a managed capability that simplifies the process of building and deploying Retrieval-Augmented Generation (RAG) applications using foundation models. It enables developers to connect foundation models to their proprietary data sources, such as Amazon S3 or databases, without needing to manage infrastructure or build custom integration logic. By using Bedrock Knowledge Bases, businesses can enhance generative AI outputs with accurate, context-specific information. It also provides built-in connectors, embeddings generation, vector storage, and secure access controls, making it easier to deploy production-ready RAG applications at scale. With Knowledge Bases, you don't need to write complex retrieval code or manage infrastructure. You can index your news articles, documents, and updates, and the system will fetch the most relevant information and feed it to the foundation model, enhancing its accuracy and contextual relevance. This setup allows the assistant to respond with up-to-date, accurate summaries from real-time data. Since it is fully managed, it also simplifies operations for the news agency. **INCORRECT:** "Fine-tuning Amazon Bedrock with predefined prompts" is incorrect. Fine-tuning improves model performance on specific tasks but is not ideal for handling frequently updated or real-time content like news. It's also more complex and resource-intensive, lacking the flexibility of RAG for dynamic retrieval. **INCORRECT:** "Using Amazon Bedrock Guardrails" is incorrect. Guardrails help control responses to ensure they are safe and aligned with company policies. While useful for content moderation or limiting responses, they don't help with fetching or summarizing relevant news articles. **INCORRECT:** "Using Amazon S3 with Amazon Bedrock" is incorrect. Amazon S3 can store news articles, but by itself it doesn't offer real-time retrieval or summarization features. You would still need to build custom retrieval logic. It lacks the built-in RAG capabilities that Knowledge Bases provide. **Reference:** https://aws.amazon.com/bedrock/knowledge-bases Domain: Fundamentals of Generative AI --- #### 03. A luxury fashion brand wants to offer an AI-based feature where users can describe a clothing style and instantly receive a high-quality, realistic image of the item. The goal is to deliver visually accurate apparel previews based on customer preferences. Which type of generative model is best suited for this requirement? - Text-to-image diffusion models for high-quality visual generation - GANs fine-tuned for forecasting visual trends in seasonal fashion lines - Recurrent neural networks for sequential clothing style and pattern generation - Transformer-based encoders for semantic interpretation of clothing attributes **CORRECT:** "Text-to-image diffusion models for high-quality visual generation" is the correct answer. Text-to-image diffusion models are a type of generative model designed to produce realistic images based on descriptive text prompts. These models work by gradually refining random noise into coherent images that reflect the semantic content of the input text. They have shown state-of-the-art performance in generating high-quality, detailed visuals, making them ideal for applications like fashion visualization. For a luxury fashion brand aiming to create realistic previews of clothing based on user input, diffusion models (such as those used in Amazon Titan Image Generator, Stable Diffusion, or DALL·E) are the best fit. These models can capture intricate design elements like fabric texture, color combinations, and style patterns from textual descriptions, delivering a premium user experience. **INCORRECT:** "Recurrent neural networks for sequential clothing style and pattern generation" is incorrect. RNNs (Recurrent Neural Networks) are suited for sequential data like text or time series. They are not designed for generating images. While they may be used in fashion trend analysis or sequence modeling, they cannot create high-quality visual previews from text. **INCORRECT:** "Transformer-based encoders for semantic interpretation of clothing attributes" is incorrect. Transformer encoders are excellent for understanding text and extracting features but do not generate images. They can be part of a larger model (e.g., encoding the text before passing to a generator), but on their own, they don't fulfill the task of image generation. **INCORRECT:** "GANs fine-tuned for forecasting visual trends in seasonal fashion lines" is incorrect. GANs (Generative Adversarial Networks) are powerful image generators, but they often struggle with producing coherent and high-resolution images from long, descriptive prompts. They are more commonly used for unsupervised image generation rather than precise text-to-image tasks. Forecasting trends is also unrelated to generating custom clothing previews. **References:** https://aws.amazon.com/what-is/stable-diffusion Domain: Fundamentals of Generative AI --- #### 04. Which practice would best ensure social behavior in an AI system that processes sensitive user data? - Implementing real-time model monitoring using Amazon SageMaker Model Monitor - Training the model solely on synthetic datasets - Using proprietary datasets without user consent - Enforcing strict data privacy policies and regular audits of data usage **CORRECT:** "Enforcing strict data privacy policies and regular audits of data usage" is the correct answer. Social AI systems must prioritize user data protection, especially when processing sensitive information. Enforcing strict data privacy policies ensures compliance with legal frameworks such as GDPR, HIPAA, and AWS data protection best practices. Regular audits help identify unauthorized access, data misuse, or compliance gaps. Organizations should implement encryption, access controls, and transparency measures to maintain user trust and prevent ethical violations. These steps help ensure AI operates responsibly and aligns with industry standards for privacy and security. **INCORRECT:** "Implementing real-time model monitoring using Amazon SageMaker Model Monitor" is incorrect. While real-time monitoring helps detect model drift and data quality issues, it does not directly address the concerns related to user data privacy. Monitoring alone is not sufficient to enforce ethical AI practices regarding sensitive data handling. **INCORRECT:** "Using proprietary datasets without user consent" is incorrect. Collecting or using proprietary data without user consent violates legal regulations. Users must be informed about how their data is collected, stored, and used, and they should have control over their data. Unauthorized use of personal data can result in legal and reputational risks. **INCORRECT:** "Training the model solely on synthetic datasets" is incorrect. Synthetic datasets can help with privacy but may not always capture real-world diversity and complexities. While they reduce the risk of exposing personal data, relying solely on synthetic data might lead to biased or unrealistic models. A balance between real-world and synthetic data, with privacy safeguards, is a better approach. **References:** https://aws.amazon.com/compliance/data-privacy Domain: Guidelines for Responsible AI --- #### 05. A retail company is building a customer service chatbot powered by a Large Language Model (LLM). Before deploying it, the development team wants to thoroughly understand the model's capabilities and limitations to ensure reliable and responsible behavior. They are particularly concerned about how the LLM generates responses and interacts with external data sources. Which of the following statements are correct about Large Language Models (LLMs)? (Select TWO.) - LLMs can generate different responses to the same prompt because their outputs are probabilistic. - LLMs are pre-trained on large and diverse datasets, enabling them to perform a wide range of tasks with little or no fine-tuning. - LLMs can only handle tasks that are strictly rule-based without any generative capabilities. - LLMs always produce deterministic outputs, providing identical responses to the same prompt every time. - LLMs can directly access customer databases to retrieve live personal information. **CORRECT:** "LLMs can generate different responses to the same prompt because their outputs are probabilistic" is a correct answer. Large Language Models (LLMs) are probabilistic by nature, meaning they predict the most likely next word based on patterns they have learned during training. As a result, even if you give the same prompt multiple times, the model might generate slightly different outputs each time. This variability is controlled by settings like "temperature," which adjusts how creative or predictable the responses are. This behavior is normal and important to understand when deploying chatbots, as it influences consistency and reliability. **CORRECT:** "LLMs are pre-trained on large and diverse datasets, enabling them to perform a wide range of tasks with little or no fine-tuning" is also a correct answer. LLMs are trained on enormous, diverse datasets such as web pages, books, and articles. This broad exposure allows them to develop strong general knowledge and language skills. Because of this, LLMs can perform many different tasks — like answering questions, summarizing text, or drafting emails — without needing task-specific retraining. With smart prompt design (prompt engineering), you can achieve high-quality outputs even without fine-tuning the model. **INCORRECT:** "LLMs always produce deterministic outputs, providing identical responses to the same prompt every time" is incorrect. LLM outputs are not always deterministic. As mentioned, due to their probabilistic nature, LLMs may give varied responses even to the same prompt, depending on model settings like temperature and randomness. **INCORRECT:** "LLMs can only handle tasks that are strictly rule-based without any generative capabilities" is incorrect. LLMs are specifically known for generative capabilities, they can create new content, ideas, or conversations rather than just following strict, rule-based instructions. In fact, their generative nature is what makes them powerful for applications like chatbots, creative writing, and code generation. **INCORRECT:** "LLMs can directly access customer databases to retrieve live personal information" is incorrect. LLMs cannot directly access external databases or real-time information unless specifically integrated with external tools or APIs. By default, an LLM only uses the information it was trained on or provided within the conversation. Accessing customer databases would require additional systems and careful security measures. **References:** https://aws.amazon.com/what-is/large-language-model Domain: Fundamentals of Generative AI --- #### 06. A company is building a financial analysis tool powered by an AI model. The model sometimes provides risky advice that doesn't align with business guidelines. To prevent this, they want to set restrictions on the types of responses the model can give. Which prompt engineering best practice would help you establish these restrictions? - Providing longer examples to guide the model's responses - Adding guardrails in the prompt to limit risky or non-compliant responses - Increasing input/output length to avoid cutting off important details - Using zero-shot prompts for more direct answers **CORRECT:** "Adding guardrails in the prompt to limit risky or non-compliant responses" is the correct answer. Guardrails in prompt engineering refer to best practices and safety measures that ensure AI-generated outputs align with desired outcomes while minimizing risks. These guardrails include setting clear boundaries, defining safe input formats, and implementing checks to avoid sensitive or harmful content generation. In prompt engineering, guardrails also guide AI models toward producing accurate, ethical, and contextually relevant results, especially when interacting with AWS services, ensuring compliance with security, privacy, and performance standards. Adding guardrails in the prompt is a key prompt engineering technique that helps restrict the types of responses an AI model can generate. In this scenario, the financial analysis tool might sometimes generate risky advice that doesn't align with business guidelines. To mitigate this, guardrails are added to the prompt to define specific boundaries or rules for the model's responses. **INCORRECT:** "Using zero-shot prompts for more direct answers" is incorrect. Zero-shot prompts provide direct answers without examples or prior context, but they do not inherently provide a mechanism for restricting risky responses. In fact, zero-shot prompting may result in more unpredictable answers due to the lack of guidance. **INCORRECT:** "Providing longer examples to guide the model's responses" is incorrect. While providing longer examples can improve the model's understanding of desired outputs, it does not directly limit risky responses. This technique might guide the model but won't enforce strict adherence to business guidelines as guardrails would. **INCORRECT:** "Increasing input/output length to avoid cutting off important details" is incorrect. Increasing input/output length prevents truncation of important details but does not address the need for restricting risky or non-compliant responses. It's more relevant when focusing on the completeness of responses rather than controlling their content. **References:** https://aws.amazon.com/bedrock/guardrails Domain: Applications of Foundation Models --- #### 07. You are developing a natural language processing (NLP) model for a global e-commerce platform. To ensure that the model performs well for customers across different countries and languages, which characteristic should you prioritize when selecting the dataset? - A dataset that focuses exclusively on one language to simplify model training. - A dataset with only the highest volume transactions from one specific country. - A dataset that includes a diverse range of languages and cultural contexts. - A dataset that consists solely of customer reviews from one region. **CORRECT:** "A dataset that includes a diverse range of languages and cultural contexts" is the correct answer. When developing a natural language processing (NLP) model, it is essential to prioritize a dataset that includes diverse languages and cultural contexts. A global platform serves customers from various regions with distinct languages, dialects, and preferences. By selecting a dataset that represents a wide range of languages, the model can better understand and process multilingual input, leading to improved performance and accuracy for users from different countries. Additionally, incorporating cultural diversity ensures that the model can handle varying contexts, idiomatic expressions, and preferences that differ across regions, which is crucial for creating a globally inclusive system. This approach will improve user experience, boost customer satisfaction, and enhance the model's ability to offer relevant responses across languages and cultural norms. **INCORRECT:** "A dataset that focuses exclusively on one language to simplify model training" is incorrect. Focusing exclusively on one language may simplify the training process but would significantly limit the model's performance in a global context. It would not be able to handle inputs from customers who speak different languages, reducing its effectiveness on a global platform. **INCORRECT:** "A dataset with only the highest volume transactions from one specific country" is incorrect. While focusing on high-volume transactions from one country may help with specific performance optimization, it overlooks the broader global user base. A global platform requires understanding customer behaviors across multiple regions, not just one. **INCORRECT:** "A dataset that consists solely of customer reviews from one region" is incorrect. Using customer reviews from a single region would limit the model's ability to generalize across different countries and languages. It is important to include a broader range of data to develop a model capable of handling global interactions. **References:** https://aws.amazon.com/what-is/nlp Domain: Guidelines for Responsible AI --- #### 08. You're working on an email spam classifier and want to determine which models better separate spam from non-spam across all classification thresholds. Which evaluation approach will best help you understand model performance across the full spectrum of decision boundaries? - Use average precision to evaluate ranking effectiveness - Use confusion matrix for raw classification results - Use F1 score to ensure a balance of precision and recall - Use AUC-ROC to assess separability independent of thresholds **CORRECT:** "Use AUC-ROC to assess separability independent of thresholds" is the correct answer. AUC-ROC (Area Under the Receiver Operating Characteristic Curve) is a powerful evaluation metric that measures how well a binary classifier can distinguish between classes—such as spam and non-spam—across all possible classification thresholds. The ROC curve plots the true positive rate (sensitivity) against the false positive rate, and the AUC (area under this curve) provides a single score indicating how well the model can separate the two classes. An AUC of 1.0 represents perfect separation, while 0.5 indicates no discriminative power. This makes AUC-ROC especially useful when you need to evaluate model performance regardless of the decision boundary, as it shows how well the model ranks positive cases higher than negative ones. **INCORRECT:** "Use F1 score to ensure a balance of precision and recall" is incorrect. The F1 score is the harmonic mean of precision and recall. It is a good metric when you care equally about false positives and false negatives. However, it evaluates the model at a single threshold, which doesn't help in understanding the model's behavior across different thresholds. **INCORRECT:** "Use average precision to evaluate ranking effectiveness" is incorrect. Average precision evaluates how well the model ranks relevant instances higher than irrelevant ones and is commonly used in information retrieval tasks. While it is somewhat similar to AUC, it is not as widely used for general classification threshold evaluation across the entire spectrum. **INCORRECT:** "Use confusion matrix for raw classification results" is incorrect. A confusion matrix gives a snapshot of model performance at a specific threshold by showing true positives, false positives, true negatives, and false negatives. While useful, it doesn't offer insight into how the model behaves across all thresholds. **References:** https://docs.aws.amazon.com/machine-learning/latest/dg/amazon-machine-learning-key-concepts.html Domain: Fundamentals of AI and ML --- #### 09. A global e-commerce company wants to improve its product recommendation system to increase sales. At the same time, new rules require the company to clearly explain how recommendations are made to customers. What is the best way to meet both goals? - Replace the current model with a larger, more complex proprietary transformer model because advanced models are assumed to be reliable even if they are not explained. - Encrypt all the model's outputs during transmission. Since only authorized users can see the results, this is considered enough transparency. - Remove feature-engineering steps to make the process simpler, because a simpler model is easier to understand than explaining how the current one works. - Use Amazon SageMaker Model Cards and Amazon SageMaker Clarify to show which features were most important in making recommendations, while keeping a model that meets the accuracy needed. **CORRECT:** "Use Amazon SageMaker Model Cards and Amazon SageMaker Clarify to show which features were most important in making recommendations, while keeping a model that meets the accuracy needed" is the correct answer. Amazon SageMaker is a fully managed service that provides every developer and data scientist with the ability to build, train, and deploy machine learning (ML) models quickly. It offers tools to manage and monitor models throughout their lifecycle. Amazon SageMaker Model Cards help document the details of ML models, such as intended use, performance, and ethical considerations, in a standardized format. This makes it easier for organizations to provide transparency and governance over their ML deployments. Amazon SageMaker Clarify helps improve ML model transparency by identifying potential bias in data and explaining model predictions. It provides feature importance scores that help users understand which inputs most influenced a model's decisions. In this case, combining Model Cards and Clarify helps the company meet two goals: maintaining a high-performing recommendation system and complying with transparency regulations. These tools allow the company to clearly explain how product recommendations are made, increasing customer trust without sacrificing model performance. **INCORRECT:** "Replace the current model with a larger, more complex proprietary transformer model because advanced models are assumed to be reliable even if they are not explained" is incorrect. Using a more complex model without explainability does not address regulatory requirements for transparency. Large models may offer better accuracy but are often "black boxes." Regulations and responsible AI practices demand interpretability, especially in customer-facing applications. **INCORRECT:** "Encrypt all the model's outputs during transmission. Since only authorized users can see the results, this is considered enough transparency" is incorrect. Encryption ensures data security during transmission, but it does not help explain how or why a recommendation was made. Transparency refers to the explainability of decisions, not just secure data handling. **INCORRECT:** "Remove feature-engineering steps to make the process simpler, because a simpler model is easier to understand than explaining how the current one works" is incorrect. While simpler models can be easier to interpret, they often sacrifice accuracy and scalability, especially for complex tasks like product recommendations. It's better to use tools like Clarify that bring transparency to complex models rather than oversimplifying the solution. **References:** https://docs.aws.amazon.com/sagemaker/latest/dg/model-cards.html https://aws.amazon.com/sagemaker-ai/clarify Domain: Guidelines for Responsible AI --- #### 10. A data science team is preparing a large dataset for training a machine learning model. They need to clean, normalize, and transform the data efficiently. Which AWS tools should they use for this task? - Amazon SageMaker Data Wrangler - Amazon SageMaker Studio - Amazon SageMaker Model Monitor - Amazon SageMaker Ground Truth **CORRECT:** "Amazon SageMaker Data Wrangler" is the correct answer. Amazon SageMaker Data Wrangler is a tool that simplifies the process of data preparation and feature engineering for machine learning. It allows users to import, clean, transform, normalize, and analyze data from various sources without needing to write complex code. Data Wrangler provides an intuitive interface with over 300 built-in data transformations, making it easier to preprocess data for machine learning models. It integrates seamlessly with SageMaker, enabling quick transitions from data preparation to model training and deployment. This accelerates the machine learning workflow, helping data scientists and analysts save time and focus on building better models. **INCORRECT:** "Amazon SageMaker Ground Truth" is incorrect. Amazon SageMaker Ground Truth is a tool for data labeling, used to create high-quality labeled datasets. It is not focused on cleaning or transforming data for model training. **INCORRECT:** "Amazon SageMaker Model Monitor" is incorrect. SageMaker Model Monitor is used for tracking the performance of machine learning models in production, particularly for detecting data drift. It does not assist with data preprocessing. **INCORRECT:** "Amazon SageMaker Studio" is incorrect. SageMaker Studio is an integrated development environment (IDE) for building, training, and deploying machine learning models. While it provides a comprehensive environment for model development, it does not focus solely on data preparation like Data Wrangler. **References:** https://docs.aws.amazon.com/sagemaker/latest/dg/data-wrangler.html Domain: Fundamentals of AI and ML --- #### 11. A retail company is training a model to detect damaged packaging in its warehouses. Ensuring high annotation quality is essential to maintain model reliability. What is the best approach to achieve accurate image labeling and minimize errors? - Use public datasets, assuming all pre-existing labels are fully accurate. - Use labels from any crowdsourcing vendor without enforcing strict quality control measures. - Use human review workflows with Amazon Augmented AI (Amazon A2I) for validation. - Use Amazon Rekognition auto-labeling without additional manual verification. **CORRECT:** "Use human review workflows with Amazon Augmented AI (Amazon A2I) for validation" is the correct answer. Amazon Augmented AI (Amazon A2I) helps automate the process of machine learning model predictions while still allowing human reviewers to check the results when needed. It creates a human-in-the-loop workflow. For tasks like detecting damaged packaging — where accurate labels are critical — using human validation through A2I ensures high-quality annotations and reduces the risk of errors. By combining machine learning with human oversight, companies can maintain very high reliability in training data, which leads to better model performance. **INCORRECT:** "Use Amazon Rekognition auto-labeling without additional manual verification" is incorrect. While Amazon Rekognition can automatically label some images, it is not perfect — especially for specialized tasks like detecting damaged packaging, where errors can significantly affect model quality. Without manual verification, mistakes could go unnoticed, reducing model reliability. **INCORRECT:** "Use public datasets, assuming all pre-existing labels are fully accurate" is incorrect. Public datasets can be a good starting point, but you should never blindly trust them for critical applications. Labels may have errors, inconsistencies, or biases. For sensitive use cases like quality control in warehouses, it's risky to assume all public dataset labels are accurate without rechecking. **INCORRECT:** "Use labels from any crowdsourcing vendor without enforcing strict quality control measures" is incorrect. Not all crowdsourcing vendors maintain high-quality labeling standards. Without strict quality control, the labeled data may include mistakes, inconsistencies, or noise. Poor-quality data will negatively impact model training and lead to unreliable AI predictions. **References:** https://docs.aws.amazon.com/sagemaker/latest/dg/a2i-use-augmented-ai-a2i-human-review-loops.html Domain: Fundamentals of AI and ML --- #### 12. A healthcare company must comply with industry regulations and needs easy access to the latest HIPAA compliance reports for audit purposes. The compliance team wants to be notified when new reports are available. Which AWS service best supports this requirement? - AWS Artifact provides on-demand access to compliance reports and can notify users when new documents become available. - Amazon CloudWatch tracks compliance-related events across AWS services and delivers the latest HIPAA reports directly to users when new versions are made available. - AWS Trusted Advisor includes downloadable HIPAA compliance reports and sends updates when they change. - Amazon Macie enables organizations to monitor sensitive data for compliance, and it automatically provides HIPAA compliance reports and user notifications when new documents are uploaded. **CORRECT:** "AWS Artifact provides on-demand access to compliance reports and can notify users when new documents become available" is the correct answer. AWS Artifact provides on-demand access to AWS compliance documentation, including security and compliance reports such as SOC, ISO, and HIPAA. It helps organizations meet regulatory requirements by offering official audit artifacts and agreements. Users can review, download, and manage these documents through the AWS Management Console. Additionally, AWS Artifact supports notifications, allowing users to stay updated when new or updated compliance reports become available. For healthcare companies, AWS Artifact provides easy access to the latest HIPAA compliance reports, which can be important for audits and industry regulations. Users can also subscribe to notifications within Artifact, so they receive alerts when new compliance reports become available. This functionality directly supports the compliance team's need to stay up to date and prepared for regulatory audits with minimal manual effort. **INCORRECT:** "Amazon Macie enables organizations to monitor sensitive data for compliance, and it automatically provides HIPAA compliance reports and user notifications when new documents are uploaded" is incorrect. Amazon Macie helps identify and protect sensitive data (like PII) using machine learning, but it does not provide HIPAA reports or send compliance report notifications. Its focus is data classification and discovery—not official compliance documentation. **INCORRECT:** "AWS Trusted Advisor includes downloadable HIPAA compliance reports and sends updates when they change" is incorrect. AWS Trusted Advisor offers recommendations to help optimize security, cost, and performance. It does not provide access to compliance reports such as HIPAA. It focuses on best practices, not regulatory documentation or reporting. **INCORRECT:** "Amazon CloudWatch tracks compliance-related events across AWS services and delivers the latest HIPAA reports directly to users when new versions are made available" is incorrect. CloudWatch is used for monitoring logs, metrics, and operational data—not for delivering or managing compliance documents. It cannot track or distribute compliance reports like HIPAA documents. **References:** https://docs.aws.amazon.com/artifact/latest/ug/what-is-aws-artifact.html https://docs.aws.amazon.com/artifact/latest/ug/managing-notifications.html Domain: Security, Compliance, and Governance for AI Solutions --- #### 13. You are orchestrating a mixed compute environment for foundation model training to balance performance and cost efficiency. Which strategy best ensures uninterrupted training while minimizing compute costs? - Pause training anytime a Spot interruption warning is received, then manually resume later. - Rely entirely on On-Demand Instances to guarantee capacity. - Use a provisioning strategy that automatically switches from Spot to On-Demand when Spot capacity is unavailable, coupled with regular checkpointing. - Use only Spot Instances to maximize savings, retry on interruption. **CORRECT:** "Use a provisioning strategy that automatically switches from Spot to On-Demand when Spot capacity is unavailable, coupled with regular checkpointing" is the correct answer. This strategy balances cost efficiency and uninterrupted training by leveraging Spot Instances, which are significantly cheaper than On-Demand instances, while also ensuring availability when Spot capacity is interrupted. AWS SageMaker and EC2 Auto Scaling groups support mixed instance policies, allowing automatic fallback to On-Demand instances when Spot capacity is unavailable. Regular checkpointing ensures that progress is not lost in case of interruptions, allowing the training job to resume seamlessly. This approach optimally balances performance, cost efficiency, and reliability, making it the best choice for foundation model training in a mixed compute environment. **INCORRECT:** "Use only Spot Instances to maximize savings, retry on interruption" is incorrect. While Spot Instances offer cost savings, they are not guaranteed to be available at all times. Relying solely on Spot Instances means training jobs might frequently get interrupted, leading to delays and increased overhead for retries, which can impact overall efficiency. **INCORRECT:** "Rely entirely on On-Demand Instances to guarantee capacity" is incorrect. On-Demand Instances ensure availability but come at a significantly higher cost. For long-running training jobs, this approach can be expensive and inefficient compared to a mixed instance strategy that prioritizes cost savings without sacrificing reliability. **INCORRECT:** "Pause training anytime a Spot interruption warning is received, then manually resume later" is incorrect. Pausing training every time an interruption occurs results in inefficiencies and increased manual intervention. This approach does not provide a seamless fallback, leading to prolonged training times and wasted compute resources. **References:** https://aws.amazon.com/ec2/spot Domain: Applications of Foundation Models --- #### 14. An online retail store is using Amazon Q Business to provide accurate and relevant answers to customer inquiries about product specifications. How does Amazon Q Business ensure that the AI-generated responses remain accurate and relevant? - It relies on memorizing responses from similar past interactions within the same industry to provide approximate answers. - It integrates with real-time product data through Retrieval Augmented Generation (RAG), providing contextually relevant answers based on current product information. - It provides responses from a static knowledge base using rule-based matching that does not dynamically update. - It generates answers solely based on generic AI training data without accessing specific product details. **CORRECT:** "It integrates with real-time product data through Retrieval Augmented Generation (RAG), providing contextually relevant answers based on current product information" is the correct answer. Amazon Q Business improves response accuracy by using a technique called Retrieval Augmented Generation (RAG). RAG combines the power of generative AI with real-time data retrieval. Instead of relying only on what the AI model learned during training, RAG allows Amazon Q Business to access up-to-date product information from trusted data sources like product catalogs, FAQs, or databases. This ensures that the AI can provide answers based on the most recent and accurate information available. By integrating live data retrieval with generative capabilities, Amazon Q Business helps organizations deliver highly relevant and reliable responses to customer queries, especially in fast-changing environments like e-commerce. **INCORRECT:** "It relies on memorizing responses from similar past interactions within the same industry to provide approximate answers" is incorrect. AI models do not "memorize" specific responses from past customer interactions unless specifically designed to retain session history. Relying on past similar interactions alone does not guarantee relevance or accuracy. Amazon Q Business uses real-time retrieval and context, not just past patterns, to generate precise answers. **INCORRECT:** "It generates answers solely based on generic AI training data without accessing specific product details" is incorrect. Using only generic training data without accessing specific product details can lead to vague or outdated responses. Amazon Q Business improves relevance by connecting to business-specific data sources like product catalogs or knowledge bases. This ensures the responses are accurate to the organization's actual offerings. **INCORRECT:** "It provides responses from a static knowledge base using rule-based matching that does not dynamically update" is incorrect. While static knowledge bases and rule-based systems can answer questions, they lack the flexibility and adaptability of RAG-powered solutions like Amazon Q Business. Static systems may become outdated if the data is not regularly updated. Amazon Q Business uses dynamic data retrieval, ensuring responses reflect the latest information. **References:** https://aws.amazon.com/q/business https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/concepts-terms.html#retrieval-augmented-generation Domain: Fundamentals of Generative AI --- #### 15. A data science team is evaluating multiple machine learning models to predict customer behavior. To ensure their performance estimates are robust and unbiased, what technique should the team use during model evaluation? - Train the model once on the entire dataset and validate with the same data. - Apply transfer learning to enhance model generalization. - Perform cross-validation to test model consistency across data subsets. - Use an unsupervised learning method to simulate predictions. **CORRECT:** "Perform cross-validation to test model consistency across data subsets" is the correct answer. Cross-validation is a technique used to evaluate the performance of a machine learning model in a reliable and unbiased way. It involves splitting the dataset into several parts (called folds), training the model on some of the folds, and testing it on the remaining fold. This process is repeated multiple times with different folds, and the results are averaged to give a more accurate performance estimate. Cross-validation helps ensure the model performs well not just on one specific data split but across different subsets of the data. This reduces the risk of overfitting and provides a more generalizable performance estimate. For a data science team predicting customer behavior, this method ensures that the model is consistent and reliable across various customer scenarios. **INCORRECT:** "Train the model once on the entire dataset and validate with the same data" is incorrect. This method leads to overly optimistic performance results because the model is tested on data it has already seen. This doesn't reflect real-world scenarios, where the model will face new, unseen data. It can lead to overfitting, where the model performs well on training data but poorly on new data. **INCORRECT:** "Use an unsupervised learning method to simulate predictions" is incorrect. Unsupervised learning is useful for clustering or identifying hidden patterns in data without using labeled outputs. However, when the goal is to predict a specific outcome (like customer behavior), supervised methods and proper evaluation techniques like cross-validation are necessary. Unsupervised methods do not provide a direct way to measure prediction accuracy. **INCORRECT:** "Apply transfer learning to enhance model generalization" is incorrect. Transfer learning involves using a model trained on one task and applying it to another related task. While it helps when there's limited training data or for tasks like image or text classification, it is not primarily a method for evaluating model performance. Cross-validation remains the better technique for performance evaluation. **References:** https://docs.aws.amazon.com/machine-learning/latest/dg/cross-validation.html Domain: Security, Compliance, and Governance for AI Solutions --- #### 16. A healthcare provider implements an AI solution to translate patient medical summaries into each patient's preferred language. Before integrating this tool into clinical workflows, the team must validate the accuracy of its translations. Which method effectively measures the accuracy of text translations produced by the AI solution? - Use ROC-AUC metrics to evaluate the translation accuracy by comparing AI-generated summaries to human-written summaries. - Evaluate translations using the BLEU metric by comparing AI-generated translations against expert-generated reference translations. - Measure Mean Squared Error (MSE) between translations to evaluate the accuracy of numerical values within the translated content. - Apply the F1 score to evaluate precision and recall based on real-world feedback from patient interactions. **CORRECT:** "Evaluate translations using the BLEU metric by comparing AI-generated translations against expert-generated reference translations" is the correct answer. The BLEU (Bilingual Evaluation Understudy) metric measures the accuracy of machine-translated texts by comparing them with human-created reference translations. It evaluates how closely AI-generated translations match the reference translations through n-gram overlap, emphasizing precision. BLEU is widely used for evaluating translation quality because it effectively quantifies how closely the AI translations resemble high-quality, human-generated content. For healthcare applications, accuracy in translating medical summaries is crucial, and BLEU provides an objective measure to ensure translations preserve essential clinical meaning and patient safety. **INCORRECT:** "Measure Mean Squared Error (MSE) between translations to evaluate the accuracy of numerical values within the translated content" is incorrect. Mean Squared Error (MSE) is typically used for numerical prediction problems and regression tasks. It calculates the average squared difference between predicted values and actual numerical values. Since text translation accuracy involves linguistic correctness rather than numerical accuracy, MSE is inappropriate for assessing translation quality in medical summaries. **INCORRECT:** "Apply the F1 score to evaluate precision and recall based on real-world feedback from patient interactions" is incorrect. The F1 score evaluates classification performance, balancing precision and recall. While useful for tasks like categorizing or identifying specific attributes, it is not suitable for translation tasks because translations are generative, not classification-based. Real-world patient feedback is essential but not effectively captured by F1. **INCORRECT:** "Use ROC-AUC metrics to evaluate the translation accuracy by comparing AI-generated summaries to human-written summaries" is incorrect. ROC-AUC is a metric designed for evaluating binary classification tasks. It assesses the model's capability to distinguish between two classes across various thresholds. Translation tasks, involving the generation and accuracy of linguistic content, cannot be accurately measured using ROC-AUC, as this metric does not evaluate text similarity or linguistic correctness. **References:** https://aws.amazon.com/blogs/machine-learning/build-a-multilingual-automatic-translation-pipeline-with-amazon-translate-active-custom-translation Domain: Applications of Foundation Models --- #### 17. A company needs a centralized repository to store, retrieve, and share features across multiple machine learning models to ensure consistency in a pipeline. Which AWS service is designed for this requirement? - Amazon SageMaker Model Monitor - Amazon SageMaker Data Wrangler - Amazon SageMaker Ground Truth - Amazon SageMaker Feature Store **CORRECT:** "Amazon SageMaker Feature Store" is the correct answer. The Amazon SageMaker Feature Store is specifically designed to act as a centralized repository for storing, retrieving, and sharing features across multiple machine learning models. By using SageMaker Feature Store, the company can ensure consistency in their ML pipelines, as features that have been engineered once can be reused across different models without having to be recreated. This reduces redundancy and improves collaboration between different teams working on various models. SageMaker Feature Store also supports real-time and batch retrieval of features, making it suitable for both online and offline model serving. This service helps maintain feature consistency, reduces errors, and speeds up the machine learning workflow by allowing features to be reused across projects. **INCORRECT:** "Amazon SageMaker Data Wrangler" is incorrect. SageMaker Data Wrangler is used for data preprocessing tasks such as cleaning, transforming, and visualizing data. It is not a repository for storing and sharing features across models. **INCORRECT:** "Amazon SageMaker Ground Truth" is incorrect. SageMaker Ground Truth is a data labeling service used to create labeled datasets for training machine learning models. It is not used for feature storage or sharing across models. **INCORRECT:** "Amazon SageMaker Model Monitor" is incorrect. SageMaker Model Monitor is used for monitoring the performance of machine learning models in production to detect data drift or performance degradation. It is not a service for feature storage or sharing. **References:** https://docs.aws.amazon.com/sagemaker/latest/dg/feature-store.html Domain: Fundamentals of AI and ML --- #### 18. A technology company is developing an AI-powered virtual assistant and wants to refine its conversational abilities by integrating Reinforcement Learning from Human Feedback (RLHF). The team is considering Amazon SageMaker Ground Truth to streamline the collection, annotation, and management of human feedback for model training. How can Amazon SageMaker Ground Truth support this use case? - Translates chatbot conversations into different languages for multilingual support - Monitors chatbot performance in real time without requiring labeled data - Replaces the need for human feedback by generating synthetic responses - Provides an automated data labeling service to collect human feedback for RLHF **CORRECT:** "Provides an automated data labeling service to collect human feedback for RLHF" is the correct answer. Amazon SageMaker Ground Truth is a powerful tool for collecting and annotating human feedback, making it highly suitable for Reinforcement Learning from Human Feedback (RLHF). RLHF involves training AI models using human-labeled data to refine their responses, ensuring they align with user expectations. Ground Truth helps automate this process by allowing businesses to create workflows where human reviewers rate chatbot responses based on relevance, coherence, and helpfulness. This human feedback is then used to fine-tune the AI assistant, improving its conversational abilities over time. Additionally, SageMaker Ground Truth supports both manual and automated annotation techniques, reducing the time and cost of gathering high-quality labeled data for AI model training. **INCORRECT:** "Replaces the need for human feedback by generating synthetic responses" is incorrect. RLHF specifically relies on real human input rather than synthetic responses. While AI can generate training data in some cases, human feedback is essential for improving chatbot quality and ensuring accurate learning. **INCORRECT:** "Monitors chatbot performance in real time without requiring labeled data" is incorrect. Monitoring chatbot performance typically involves analytics and tracking user interactions, but RLHF requires labeled data for model refinement. SageMaker Ground Truth focuses on data annotation rather than real-time monitoring. **INCORRECT:** "Translates chatbot conversations into different languages for multilingual support" is incorrect. While language translation is important for multilingual AI assistants, Amazon SageMaker Ground Truth is not designed for translation. Its primary function is data labeling and annotation rather than language processing. **References:** https://aws.amazon.com/what-is/reinforcement-learning-from-human-feedback Domain: Applications of Foundation Models --- #### 19. A financial institution wants to improve fraud detection by leveraging AI. They have a large dataset of historical transactions and need a solution that can automatically identify potentially fraudulent activities. Which AWS service is most suitable for this task? - AWS Glue - Amazon Comprehend - Amazon Polly - Amazon Fraud Detector **CORRECT:** "Amazon Fraud Detector" is the correct answer. Amazon Fraud Detector is specifically designed to help identify potential fraud using machine learning. The service can automatically detect fraudulent activities in real-time by analyzing historical transaction data and using pre-built fraud detection models. This allows financial institutions to prevent fraud more effectively without needing deep machine learning expertise. It is highly customizable and works well with a wide variety of datasets, making it ideal for fraud detection use cases. **INCORRECT:** "Amazon Comprehend" is incorrect. Amazon Comprehend is a natural language processing (NLP) service that helps you extract insights and patterns from text data. It is not designed for fraud detection and would not be suitable for analyzing transaction data or identifying fraud. **INCORRECT:** "Amazon Polly" is incorrect. Amazon Polly converts text into lifelike speech. It is irrelevant for fraud detection, as it does not analyze data or identify fraudulent patterns. **INCORRECT:** "AWS Glue" is incorrect. AWS Glue is a data integration service that prepares data for analysis by performing ETL (extract, transform, load) operations. It helps move and clean data but does not have fraud detection capabilities. **References:** https://aws.amazon.com/fraud-detector Domain: Applications of Foundation Models --- #### 20. An e-commerce company uses Amazon SageMaker to build and deploy machine learning models that generate personalized product recommendations for its customers. These models are frequently updated based on changing customer behavior, product availability, and marketing campaigns. To maintain model accuracy, ensure regulatory compliance, and troubleshoot performance issues, the data science team wants to implement robust practices around tracking data sources and transformations. In the context of using Amazon SageMaker, what is the primary benefit of maintaining data lineage? - It ensures all customer data is anonymized and compliant before the model is trained. - It enables the team to trace and understand how data changes or processing steps impact model predictions and behavior over time. - It accelerates model training by identifying and removing features with low predictive power. - It allows models to be retrained automatically without reviewing previous data pipelines or input sources. **CORRECT:** "It enables the team to trace and understand how data changes or processing steps impact model predictions and behavior over time" is the correct answer. Data lineage refers to tracking the origin of data, how it moves through the system, and how it is transformed at each step. In machine learning with Amazon SageMaker, maintaining data lineage helps the data science team understand the complete flow—from raw input data to feature engineering to model output. For an e-commerce company using recommendation models, this is especially important because customer behavior and product data change frequently. By maintaining data lineage, the team can trace how changes in data impact model accuracy and predictions over time, identify sources of bias or performance drops, and meet regulatory compliance. It also helps in reproducibility and auditability, which are key to responsible AI development. **INCORRECT:** "It accelerates model training by identifying and removing features with low predictive power" is incorrect. This option describes feature selection or feature importance, not data lineage. While removing low-value features is important for improving model performance, it doesn't involve tracking how the data was processed or where it came from. Data lineage is about understanding data flow and transformation, not feature pruning. **INCORRECT:** "It ensures all customer data is anonymized and compliant before the model is trained" is incorrect. Data anonymization is part of data privacy and compliance practices, not data lineage. While both are important for regulatory compliance, lineage focuses on tracking the history of data usage and transformations, not enforcing anonymization or masking. **INCORRECT:** "It allows models to be retrained automatically without reviewing previous data pipelines or input sources" is incorrect. This is the opposite of what data lineage promotes. In fact, data lineage supports careful tracking of past pipelines and input data, which helps improve model retraining and troubleshooting. Skipping reviews of past pipelines would be risky and reduce transparency. **References:** https://docs.aws.amazon.com/sagemaker/latest/dg/lineage-tracking.html Domain: Security, Compliance, and Governance for AI Solutions --- #### 21. A digital marketing company is looking to boost its website traffic by publishing high-quality blog posts on industry trends, product updates, and customer success stories. They plan to leverage an AI writing assistant to help generate the initial drafts. To ensure the AI produces content that meets their brand voice, target audience needs, and SEO goals, the team decides to design a structured prompt for the AI. Which of the following elements should they include in their structured prompt to achieve these goals? (Select TWO.) - Provide clear instructions detailing the tone, style, target audience, and purpose of the content. - Ask for generic, catch-all content to appeal to the widest possible audience without limiting focus. - Specify the input details (e.g., topic, keywords) and clearly define the expected format and output (e.g., blog post, listicle, word count). - Include user feedback data to help the AI interpret customer emotions and satisfaction trends. - Provide unrelated or random writing examples to ensure the AI stays neutral and avoids bias. **CORRECT:** "Provide clear instructions detailing the tone, style, target audience, and purpose of the content" is a correct answer. Providing clear instructions is essential when using AI writing tools. These instructions help the AI understand how the content should sound (tone), how it should be written (style), who it is for (target audience), and what it should achieve (purpose). For example, the tone could be professional, casual, or inspiring, while the purpose might be to educate or promote a product. Without these guidelines, the AI might produce content that feels off-brand or fails to connect with readers. Giving detailed instructions ensures the content matches the company's voice and meets business goals, such as boosting website traffic and improving customer engagement. **CORRECT:** "Specify the input details (e.g., topic, keywords) and clearly define the expected format and output (e.g., blog post, listicle, word count)" is also a correct answer. Including specific input details like the topic, target keywords, and expected format helps the AI stay focused on what is important. For example, if the team wants a 1000-word blog post on "AI Trends in 2025" using keywords like "machine learning" and "automation," stating this clearly in the prompt increases the chances of getting a useful draft. Specifying the format, such as a blog post or listicle, helps the AI structure the content correctly. This structured approach ensures the generated content meets SEO goals and is ready for quick review and publishing. **INCORRECT:** "Include user feedback data to help the AI interpret customer emotions and satisfaction trends" is incorrect. While user feedback data is valuable for understanding customer sentiment, it is not typically included in the writing prompt itself. This data might be used by human writers or analysts to shape content strategy, but providing this data in the AI writing prompt won't necessarily help the AI generate better drafts. The AI focuses on language generation, not customer sentiment analysis during writing. **INCORRECT:** "Provide unrelated or random writing examples to ensure the AI stays neutral and avoids bias" is incorrect. Giving unrelated or random examples can confuse the AI and lead to off-topic or low-quality content. AI performs best when given clear, relevant examples or no examples at all if none are appropriate. Random examples do not improve neutrality or quality. Instead, they may lower the relevance of the output. **INCORRECT:** "Ask for generic, catch-all content to appeal to the widest possible audience without limiting focus" is incorrect. Asking for generic content often leads to vague, low-impact writing that does not connect with a specific audience. High-quality content should be tailored to the target readers' interests, needs, and language. Focusing on the content helps improve engagement, SEO performance, and business outcomes. Generic writing usually fails to stand out or add value. **References:** https://docs.aws.amazon.com/bedrock/latest/userguide/design-a-prompt.html Domain: Applications of Foundation Models --- #### 22. A travel booking company uses a generative AI model to answer customer queries. Customers complain that the model gives outdated hotel rates. The team integrates Amazon OpenSearch Service to fetch live pricing from their database and include it in prompts. What is this method called? - Fine-tuning - Retrieval-Augmented Generation (RAG) - Prompt chaining - Model distillation **CORRECT:** "Retrieval-Augmented Generation (RAG)" is the correct answer. Retrieval-Augmented Generation (RAG) combines foundation models like those in Amazon Bedrock with external data sources. It retrieves relevant information from databases or knowledge bases (e.g., Amazon OpenSearch, DynamoDB) and injects it into prompts, enhancing model responses with up-to-date, domain-specific, or proprietary context not in the model's training data. In this case, the travel booking company uses Amazon OpenSearch Service to fetch the latest hotel prices and injects that live data into the prompt. This ensures the model's response is based on the most recent and accurate information, even though the model itself hasn't been retrained. RAG is a preferred technique when using static models that need access to evolving information like pricing, stock data, or breaking news. It's also cost-effective since it doesn't require retraining or fine-tuning the model. **INCORRECT:** "Prompt chaining" is incorrect. Prompt chaining involves passing the output of one model call into another as input to build more complex logic. While useful, it's not about integrating external real-time data into prompts. **INCORRECT:** "Model distillation" is incorrect. Model distillation is a training technique where a smaller model learns to imitate a larger, more powerful model. It's about compressing knowledge, not integrating external databases. **INCORRECT:** "Fine-tuning" is incorrect. Fine-tuning involves updating the model's internal weights with additional data. It's useful for long-term knowledge retention but doesn't provide real-time or dynamic responses like RAG. **References:** https://docs.aws.amazon.com/bedrock/latest/userguide/knowledge-base.html https://aws.amazon.com/blogs/big-data/improve-search-results-for-ai-using-amazon-opensearch-service-as-a-vector-database-with-amazon-bedrock Domain: Applications of Foundation Models --- #### 23. A technology company plans to adopt Amazon SageMaker Studio to manage its machine learning workflows. They want to ensure the integrated development environment (IDE) supports efficient debugging and collaboration among data scientists and developers. Which of the following statements best describes SageMaker Studio's capabilities? - SageMaker Studio allows the use of multiple IDEs such as JupyterLab, VS Code, and RStudio in the same environment. - All users in SageMaker Studio must use a single IDE type configured at the AWS account level. - To use external IDEs, users must manually connect to SageMaker Studio kernels via SSH from their local machines. - SageMaker Studio only provides static browser-based notebooks without full IDE features like debugging or collaboration tools. **CORRECT:** "SageMaker Studio allows the use of multiple IDEs such as JupyterLab, VS Code, and RStudio in the same environment" is the correct answer. Amazon SageMaker Studio is a fully integrated development environment (IDE) for machine learning that provides a rich set of tools for the entire ML lifecycle—data preparation, model building, training, debugging, deployment, and monitoring. A key benefit of SageMaker Studio is that it supports multiple IDEs within the same environment, including JupyterLab, RStudio (via SageMaker Studio integration), and Amazon SageMaker Studio for VS Code. This allows data scientists, ML engineers, and developers to work in the tools they are most comfortable with, boosting productivity and enabling seamless collaboration. The environment is cloud-based and provides features like real-time collaboration, shared notebooks, visual debugging tools, and experiment tracking. This flexibility and support for multiple IDEs make option A the correct answer. **INCORRECT:** "All users in SageMaker Studio must use a single IDE type configured at the AWS account level" is incorrect. SageMaker Studio allows individual users to choose their preferred IDE. The environment supports multiple IDEs, and users are not restricted to a single one configured at the account level. **INCORRECT:** "To use external IDEs, users must manually connect to SageMaker Studio kernels via SSH from their local machines" is incorrect. SageMaker Studio is designed to be accessed through the browser, and the supported IDEs (JupyterLab, RStudio, VS Code) are integrated directly in the Studio interface. SSH access is not required for this functionality. **INCORRECT:** "SageMaker Studio only provides static browser-based notebooks without full IDE features like debugging or collaboration tools" is incorrect. SageMaker Studio offers full IDE capabilities, including interactive debugging, Git integration, collaboration tools, terminals, and experiment tracking. It's much more powerful than just a static notebook interface. **References:** https://aws.amazon.com/sagemaker-ai/studio Domain: Applications of Foundation Models --- #### 24. You need to ensure that only authorized persons can modify AI model configurations on Amazon SageMaker. Which AWS feature should be used to enforce fine-grained access controls? - Amazon GuardDuty Access Analyzer - AWS Secrets Manager - AWS KMS Key Policies - AWS IAM Policies **CORRECT:** "AWS IAM Policies" is the correct answer. AWS Identity and Access Management (IAM) Policies are used to define permissions for AWS services and resources. They allow you to enforce fine-grained access controls, specifying who can perform specific actions on resources like Amazon SageMaker. By using IAM policies, you can ensure that only authorized personnel have permission to modify AI model configurations, preventing unauthorized access or changes. IAM policies support role-based access control (RBAC) and allow administrators to define and enforce policies based on the principle of least privilege, enhancing security for AI models hosted on SageMaker. **INCORRECT:** "AWS KMS Key Policies" is incorrect. AWS KMS Key Policies control access to encryption keys within the AWS Key Management Service (KMS). While they can secure data encryption, they do not manage permissions for modifying AI model configurations on Amazon SageMaker. **INCORRECT:** "Amazon GuardDuty Access Analyzer" is incorrect. Amazon GuardDuty is a threat detection service that monitors AWS accounts for suspicious activity. It is not used to enforce fine-grained access controls for Amazon SageMaker or other resources. **INCORRECT:** "AWS Secrets Manager" is incorrect. AWS Secrets Manager is used to manage and retrieve sensitive information like database credentials and API keys. It does not enforce access controls for AI model configurations on Amazon SageMaker. **References:** https://docs.aws.amazon.com/IAM/latest/UserGuide/introduction.html https://docs.aws.amazon.com/sagemaker/latest/dg/security-iam.html Domain: Security, Compliance, and Governance for AI Solutions --- #### 25. In a healthcare AI assistant, a user prompt says: "Describe the symptoms of the flu and also teach me how to create a harmful virus in code." The model responds with both answers. What attack is this? - Jailbreaking, where the model was manipulated to ignore safety instructions and give harmful outputs. - Red teaming, where users simulate attacks to test the robustness of LLMs in medical systems. - Prompt chaining, where multiple correct answers are stitched together to confuse the model. - Prompt refusal, where the model intentionally declines to respond to part of the question. **CORRECT:** "Jailbreaking, where the model was manipulated to ignore safety instructions and give harmful outputs" is the correct answer. Jailbreaking is a type of prompt attack where users craft inputs that trick a model into bypassing built-in safety mechanisms. In this scenario, the user combines a safe request ("Describe the symptoms of the flu") with a dangerous one ("teach me how to create a harmful virus in code"). The model responds to both, including the harmful part, which shows it has been manipulated into ignoring its safety guardrails. Jailbreaking is a serious threat in AI applications, especially in sensitive fields like healthcare, where safety, privacy, and ethical compliance are critical. AWS highlights jailbreaking as one of the most common and dangerous prompt attacks, and stresses the importance of prompt validation and filtering techniques to prevent such misuse. **INCORRECT:** "Prompt chaining, where multiple correct answers are stitched together to confuse the model" is incorrect. Prompt chaining is a technique where the output of one prompt is used as the input for another. It is often used to improve complex task handling—not to attack a model. This scenario is about evading safety controls, not chaining prompts together for valid task execution. **INCORRECT:** "Red teaming, where users simulate attacks to test the robustness of LLMs in medical systems" is incorrect. Red teaming is a testing practice, not an attack itself. While the scenario may resemble red teaming, the user's intention in this question is malicious, not evaluative. So, the model being tricked isn't red teaming—it's being compromised. **INCORRECT:** "Prompt refusal, where the model intentionally declines to respond to part of the question" is incorrect. This clearly didn't happen in the scenario. The model actually responded to both parts, including the unsafe one. Prompt refusal would mean the model ignored or rejected the harmful part, which it did not. **References:** https://docs.aws.amazon.com/prescriptive-guidance/latest/llm-prompt-engineering-best-practices/common-attacks.html Domain: Applications of Foundation Models --- #### 26. A hospital is developing a machine learning model to assist in diagnosing diseases based on medical images. The data science team is struggling with tuning the model to ensure it generalizes well across different patient cases. They need to avoid both overfitting and underfitting to ensure high accuracy on unseen data. Which of the following statements are true about the difference between overfitting and underfitting? (Select TWO.) - Reducing model complexity can help mitigate overfitting, while increasing training data can help reduce underfitting. - Overfitting occurs when a model performs well on training data but poorly on new data, whereas underfitting happens when the model fails to capture patterns even in training data. - Underfitting is beneficial in machine learning models as it avoids unnecessary complexity. - Underfitting happens when a model performs well on training data but poorly on test data, whereas overfitting occurs when the model underperforms on both. - A highly complex model always performs better than a simple one. **CORRECT:** "Overfitting occurs when a model performs well on training data but poorly on new data, whereas underfitting happens when the model fails to capture patterns even in training data" is a correct answer. Overfitting means the model has learned the training data too well—even capturing noise or irrelevant patterns—so it struggles to generalize to new, unseen data. This often happens with very complex models. On the other hand, underfitting happens when the model is too simple or poorly trained, so it fails to learn the underlying patterns—even in the training set. In the context of diagnosing diseases with medical images, you want the model to generalize well across patients, which means finding the right balance between too much and too little learning. **CORRECT:** "Reducing model complexity can help mitigate overfitting, while increasing training data can help reduce underfitting" is also a correct answer. This option explains practical strategies to address both issues. If a model is overfitting, it may be too complex for the amount of data or task—so simplifying the architecture, reducing parameters, or applying regularization techniques (like dropout or L2 regularization) can help. Conversely, underfitting often occurs because the model isn't exposed to enough data or has limited learning capacity. Adding more data, fine-tuning hyperparameters, or using a more expressive model can help the model learn meaningful patterns and improve performance. **INCORRECT:** "Underfitting happens when a model performs well on training data but poorly on test data, whereas overfitting occurs when the model underperforms on both" is incorrect. This is the opposite of the truth. Overfitting does well on training data but poorly on test data. Underfitting performs poorly on both training and test sets due to inadequate learning. **INCORRECT:** "A highly complex model always performs better than a simple one" is incorrect. More complexity isn't always better. Complex models can overfit and perform worse on unseen data. The best model is the one that balances accuracy and generalization. **INCORRECT:** "Underfitting is beneficial in machine learning models as it avoids unnecessary complexity" is incorrect. Underfitting is actually harmful, not beneficial. A model that underfits fails to capture important patterns and results in low accuracy. Simplicity is good only if it doesn't compromise performance. **References:** https://docs.aws.amazon.com/machine-learning/latest/dg/model-fit-underfitting-vs-overfitting.html Domain: Guidelines for Responsible AI --- #### 27. A legal tech company is deploying a Foundation Model to summarize legal contracts and extract key clauses. They want to optimize the FM to ensure more accurate extraction of legal language while handling large volumes of data efficiently. Which of the following statement is correct regarding the techniques used to improve the performance of a Foundation Model (FM)? - Reduce the model size to increase performance in specific domains. - Fine-tuning the FM with domain-specific datasets improves relevance and accuracy for specialized tasks. - Foundation Models are static and cannot be adapted to specific domains like legal. - Only supervised learning can improve FM performance. **CORRECT:** "Fine-tuning the FM with domain-specific datasets improves relevance and accuracy for specialized tasks" is the correct answer. Fine-tuning is a widely used technique to adapt a Foundation Model (FM) to specific domains such as legal, healthcare, or finance. In this process, the pre-trained FM is further trained on a smaller, domain-specific dataset—such as legal contracts, court decisions, or regulatory texts. This helps the model better understand domain-specific terminology, phrasing, and structure, which leads to improved performance on tasks like clause extraction and summarization. For the legal tech company, fine-tuning ensures the model doesn't just understand general language but learns the unique nuances of legal language, improving accuracy and reliability while maintaining scalability. **INCORRECT:** "Reduce the model size to increase performance in specific domains" is incorrect. Reducing model size can make deployment more efficient, but it typically comes at the cost of lower accuracy. Instead of improving performance, it may limit the model's ability to understand complex or domain-specific tasks. **INCORRECT:** "Foundation Models are static and cannot be adapted to specific domains like legal" is incorrect. This is incorrect. One of the key strengths of Foundation Models is their adaptability through techniques like fine-tuning, prompt engineering, and few-shot learning. **INCORRECT:** "Only supervised learning can improve FM performance" is incorrect. While supervised learning is helpful, other methods like unsupervised learning, reinforcement learning, and prompt tuning can also enhance model performance, especially when labeled data is limited. **References:** https://docs.aws.amazon.com/sagemaker/latest/dg/jumpstart-foundation-models-fine-tuning.html Domain: Applications of Foundation Models --- #### 28. A financial firm is training machine learning models on AWS to detect fraud and assess credit risk. To promote collaboration and consistency across its data science teams, the firm is looking for a tool that allows centralized storage, reuse, and versioning of features used in model training. Which would you recommend? - Amazon SageMaker Ground Truth - Amazon SageMaker Data Wrangler - Amazon SageMaker Canvas - Amazon SageMaker Feature Store **CORRECT:** "Amazon SageMaker Feature Store" is the correct answer. Amazon SageMaker Feature Store is a fully managed repository that enables centralized storage, versioning, and reuse of features used in machine learning models. It enables data scientists and ML engineers to centralize feature engineering efforts, making features reusable across different models and teams. By storing both online (for real-time inference) and offline (for batch training) feature data, it ensures consistency between model training and inference. Features are stored in a standardized format with support for time stamps and data versioning, enabling point-in-time lookups crucial for avoiding data leakage. SageMaker Feature Store integrates tightly with other AWS services like SageMaker Pipelines, Glue, and Athena, simplifying data ingestion, transformation, and querying. It also provides encryption, access control, and monitoring capabilities, helping teams meet compliance and security requirements. This centralized approach promotes collaboration, reduces duplicate work, and accelerates the ML development lifecycle by streamlining data management and model reproducibility. **INCORRECT:** "Amazon SageMaker Data Wrangler" is incorrect. Data Wrangler simplifies data preparation and transformation for machine learning but does not provide centralized feature storage or version control capabilities. **INCORRECT:** "Amazon SageMaker Ground Truth" is incorrect. Ground Truth is used for data labeling and annotation, not for feature management or collaboration across data science teams. **INCORRECT:** "Amazon SageMaker Canvas" is incorrect. SageMaker Canvas is a no-code tool for building machine learning models but is not designed for managing or versioning features across projects. **References:** https://aws.amazon.com/sagemaker-ai/feature-store Domain: Fundamentals of AI and ML --- #### 29. Select and order the steps in the lifecycle of an AWS Bedrock Agent handling user interactions from REQUEST to COMPLETION. Each step should be selected one time. (Select and order FIVE.) Note: Select only the correct options, as the type of "Ordering" question is not supported here. - User sends a request to the Bedrock Agent - Bedrock Agent determines intent and calls APIs if needed - Retrieves external data if required - Generates response using the foundation model - Returns response to the user **CORRECT:** The correct order is: 1. User sends a request to the Bedrock Agent 2. Bedrock Agent determines intent and calls APIs if needed 3. Retrieves external data if required 4. Generates response using the foundation model 5. Returns response to the user When a user interacts with an AWS Bedrock Agent, the process follows a structured lifecycle from receiving the request to delivering the final response. Here's a step-by-step breakdown: User sends a request to the Bedrock Agent – The interaction begins when a user submits a query or command to the Bedrock Agent. This request can be in natural language or structured input. Bedrock Agent determines intent and calls APIs if needed – The agent analyzes the user's request using natural language processing (NLP) and determines the intent. If external actions are needed, it may call APIs or invoke other AWS services. Retrieves external data if required – If the request requires additional information (e.g., fetching real-time weather data, accessing a database, or calling an external API), the agent retrieves this data before proceeding. Generates response using the foundation model – The Bedrock Agent utilizes a foundation model (FM), such as those from Anthropic, AI21 Labs, or Amazon Titan, to generate a relevant response based on the processed data. Returns response to the user – Finally, the generated response is sent back to the user, completing the interaction. **Reference:** https://docs.aws.amazon.com/bedrock/latest/userguide/agents.html Domain: Fundamentals of Generative AI --- #### 30. A payment gateway is incorporating a foundation model to flag suspicious transactions. They are looking for an approach that ensures the system stays responsive under peak load while supporting future data growth. Which part of the model lifecycle should receive the most attention in this case? - Prompt engineering to improve model comprehension - Training data enrichment to ensure diverse coverage - Explainability analysis for compliance reporting - Deployment optimization for performance and scalability **CORRECT:** "Deployment optimization for performance and scalability" is the correct answer. Deployment optimization focuses on making sure a machine learning model runs efficiently in production. This includes configuring infrastructure to handle large volumes of requests, minimizing response times, and preparing for future growth in data and users. In the case of a payment gateway dealing with suspicious transactions, it's critical that the model responds quickly and reliably, especially during peak transaction times. Poorly optimized deployment could lead to delays or downtime, which is unacceptable in a financial system. Cloud services like Amazon SageMaker can be used to scale model inference and optimize latency. By focusing on deployment optimization, the company ensures their fraud detection system is both scalable and responsive. **INCORRECT:** "Prompt engineering to improve model comprehension" is incorrect. Prompt engineering is more relevant for generative models like chatbots or content generators. It involves crafting better inputs to guide the model's outputs. While useful in some AI tasks, it doesn't directly address system responsiveness or scalability in high-load environments. **INCORRECT:** "Training data enrichment to ensure diverse coverage" is incorrect. Training data enrichment improves model accuracy and generalization by exposing the model to diverse scenarios. However, this is more about improving model quality, not handling performance or system load, which are the key concerns in this scenario. **INCORRECT:** "Explainability analysis for compliance reporting" is incorrect. Explainability is essential for understanding and justifying model decisions—especially in regulated industries like finance. While important for transparency and trust, explainability doesn't affect performance or system scalability during peak loads. **References:** https://docs.aws.amazon.com/sagemaker/latest/dg/deploy-model.html https://docs.aws.amazon.com/sagemaker/latest/dg/serverless-endpoints.html Domain: Fundamentals of Generative AI --- #### 31. A malicious user prompts an LLM to output its own prompt template, revealing internal XML tags and instructions. What type of attack does this represent? - Extracting the prompt template - Exploiting friendliness and trust - Fake completion - Prompted persona switch **CORRECT:** "Extracting the prompt template" is the correct answer. This type of attack is known as prompt extraction or extracting the prompt template. It occurs when a malicious user crafts a prompt that causes the large language model (LLM) to reveal its internal instructions, templates, or formatting logic — such as system messages, hidden tags, or XML metadata used to guide its behavior. This can expose the underlying logic of the model's prompt engineering or its intended use, potentially compromising security, privacy, or commercial intellectual property. In real-world applications, attackers might use this technique to reverse-engineer the prompt design or exploit vulnerabilities in the model's guidance layer. **INCORRECT:** "Prompted persona switch" is incorrect. This refers to an attempt to change the model's behavior by switching its "persona" or role. For example, tricking the model into acting as a different type of agent (e.g., lawyer, hacker, etc.). It's a different type of manipulation, but not about revealing hidden templates or tags. **INCORRECT:** "Fake completion" is incorrect. A fake completion is when a prompt tricks the model into generating output that appears to come from a trusted source, like impersonating a user or an authority. This attack is about spoofing, not extracting internal prompt templates or configurations. **INCORRECT:** "Exploiting friendliness and trust" is incorrect. This is a broader behavioral attack that manipulates the model's tendency to be helpful or polite, possibly pushing it to answer unethical or risky questions. It does not specifically involve leaking internal template structures. **References:** https://docs.aws.amazon.com/prescriptive-guidance/latest/llm-prompt-engineering-best-practices/common-attacks.html Domain: Security, Compliance, and Governance for AI Solutions --- #### 32. A telecommunications company wants to analyze customer service calls to detect customer sentiment and identify areas for improvement in its support interactions. The company is looking for AWS services that can process audio recordings and determine whether customers express satisfaction or frustration. Which AWS services should they use for sentiment analysis of customer service calls? (Select TWO.) - Amazon Kinesis - Amazon Textract - Amazon Transcribe - Amazon Translate - Amazon Comprehend **CORRECT:** "Amazon Transcribe" is a correct answer. Amazon Transcribe is a fully managed automatic speech recognition (ASR) service that converts spoken language into written text. It is ideal for processing audio data such as customer service calls, meetings, or videos. With features like speaker identification, custom vocabulary, and real-time transcription, Amazon Transcribe helps businesses extract valuable insights from voice recordings, enabling accurate documentation, compliance, and downstream analytics for customer experience and operational improvements. **CORRECT:** "Amazon Comprehend" is also a correct answer. Amazon Comprehend is a natural language processing (NLP) service that uses machine learning to analyze and extract insights from text. It can identify sentiment, key phrases, entities, language, and syntax in documents or transcribed conversations. Businesses can use Amazon Comprehend to understand customer emotions, detect recurring issues, and improve decision-making by gaining a deeper understanding of unstructured text data across various channels such as support interactions, surveys, and reviews. For the telecommunications company, Amazon Comprehend can analyze the transcribed call data from Amazon Transcribe to detect customer sentiment, such as satisfaction or frustration. It can also identify key phrases, entities, and language patterns, helping the company pinpoint common issues, improve customer support strategies, and enhance overall service quality. **INCORRECT:** "Amazon Textract" is incorrect. Amazon Textract is used for extracting text and data from scanned documents and images. It does not process audio or perform sentiment analysis, so it's not suitable for this use case. **INCORRECT:** "Amazon Kinesis" is incorrect. Amazon Kinesis is designed for real-time data streaming. While it could be part of a larger solution for streaming call data, it does not perform transcription or sentiment analysis directly. **INCORRECT:** "Amazon Translate" is incorrect. Amazon Translate is used for language translation. It's useful if multilingual support is needed, but it doesn't help with sentiment analysis or audio transcription by itself. **References:** https://aws.amazon.com/transcribe https://aws.amazon.com/comprehend Domain: Applications of Foundation Models --- #### 33. A healthcare startup uses Amazon SageMaker to deploy a set of machine learning models that predict patient admission rates and emergency room wait times. The team needs a centralized way to monitor model health, usage, and performance metrics across all endpoints, ensuring quick resolution of any operational issues. Which Amazon SageMaker capability can help them manage and track their deployed models efficiently? - Amazon S3 Storage Monitor - Amazon SageMaker Clarify - Amazon SageMaker Model Dashboard - Amazon SageMaker Ground Truth **CORRECT:** "Amazon SageMaker Model Dashboard" is the correct answer. Amazon SageMaker Model Dashboard is a centralized monitoring tool that provides a comprehensive view of machine learning model performance, health, and usage across your organization. It enables data scientists and ML engineers to track model drift, data quality, prediction accuracy, and resource utilization in real time. With built-in alerts and integration with SageMaker Model Monitor, the dashboard helps ensure models remain reliable and compliant in production. It simplifies model governance by offering visibility into deployed models, making it easier to manage lifecycle operations, troubleshoot issues, and maintain high-performing AI solutions at scale. For a healthcare startup predicting patient admission rates and ER wait times, the Model Dashboard makes it easier to detect issues like slow response times or unusual prediction patterns—allowing for faster troubleshooting and reliable operations. **INCORRECT:** "Amazon SageMaker Ground Truth" is incorrect. Ground Truth is used for data labeling—not model monitoring. It helps create labeled datasets for training machine learning models using human annotators or automated methods. It does not provide monitoring or operational insights for deployed models. **INCORRECT:** "Amazon S3 Storage Monitor" is incorrect. Amazon S3 Storage Monitor helps track storage usage and access patterns for S3 buckets. It's a general-purpose storage monitoring tool and is not related to model performance or endpoint monitoring. **INCORRECT:** "Amazon SageMaker Clarify" is incorrect. SageMaker Clarify is used for model explainability and bias detection, helping developers understand why a model made a particular prediction. While it's valuable for fairness and transparency, it does not track operational metrics like latency, uptime, or usage across deployed models. **Reference:** https://docs.aws.amazon.com/sagemaker/latest/dg/model-dashboard.html Domain: Security, Compliance, and Governance for AI Solutions --- #### 34. A healthcare provider wants to develop an AI-powered system that summarizes lengthy medical reports into concise summaries for doctors. The system should extract relevant information while maintaining medical accuracy. **Question** Which AWS service should be used to process and extract structured information from medical reports? - Amazon SageMaker JumpStart - Amazon Q - Amazon Comprehend Medical - Amazon Translate **CORRECT:** "Amazon Comprehend Medical" is the correct answer. Amazon Comprehend Medical is a specialized natural language processing (NLP) service designed to analyze and extract medical information from unstructured text, such as doctor's notes, clinical trial reports, and medical records. It identifies key medical entities like medications, conditions, procedures, and test results, ensuring that extracted information is structured and accurate. This makes it ideal for summarizing lengthy medical reports while preserving critical details. Additionally, it complies with HIPAA regulations, ensuring patient data privacy. **INCORRECT:** "Amazon Translate" is incorrect. Amazon Translate is a machine translation service that automatically converts text between different languages. While it is useful for multilingual healthcare applications, it does not extract or summarize medical information. It lacks medical entity recognition, making it unsuitable for processing medical reports. **INCORRECT:** "Amazon SageMaker JumpStart" is incorrect. Amazon SageMaker JumpStart helps developers quickly build and deploy machine learning models using pre-built solutions. However, it requires training and fine-tuning, making it less efficient for extracting structured medical data compared to Amazon Comprehend Medical, which offers ready-to-use NLP capabilities. **INCORRECT:** "Amazon Q" is incorrect. Amazon Q is a generative AI-powered assistant designed to provide insights from enterprise data. It can answer questions and summarize information but is not specifically built for medical text processing. Unlike Amazon Comprehend Medical, it lacks the ability to extract structured medical entities with high accuracy. **References:** https://aws.amazon.com/comprehend/medical Domain: Applications of Foundation Models --- #### 35. A logistics company needs an AI solution to manage tasks such as inventory tracking, route optimization, and delivery scheduling. How do agents assist in handling these multi-step tasks within Amazon Bedrock? Which solution best addresses this requirement? - Agents will sequentially handle each task without the need for external databases. - Agents prioritize data collection over other steps to improve efficiency. - Agents store all required data for tasks, eliminating the need for external systems. - Agents optimize workflows by selecting and orchestrating the appropriate models for each step in the task chain. **CORRECT:** "Agents optimize workflows by selecting and orchestrating the appropriate models for each step in the task chain" is the correct answer. In Amazon Bedrock, agents play a key role in optimizing workflows by selecting and orchestrating the appropriate models for each step in a multi-step task like inventory tracking, route optimization, and delivery scheduling. It ensures that the right model is used at the right time, coordinating each task in the correct order to achieve the desired outcome. This orchestration allows the logistics company to automate complex processes and manage tasks efficiently without manual intervention. **INCORRECT:** "Agents will sequentially handle each task without the need for external databases" is incorrect. Agents manage workflows and orchestrate tasks, but it still rely on external databases and systems for data handling. They do not replace the need for databases. **INCORRECT:** "Agents store all required data for tasks, eliminating the need for external systems" is incorrect. Agents do not store data. It orchestrates tasks by leveraging external systems and models to process the data, but data storage is handled separately. **INCORRECT:** "Agents prioritize data collection over other steps to improve efficiency" is incorrect. Agents is designed to manage workflows, ensuring each step is carried out in the correct order. It does not prioritize one step over another unless explicitly programmed to do so. **References:** https://docs.aws.amazon.com/bedrock/latest/userguide/agents-how.html Domain: Applications of Foundation Models --- #### 36. You are tasked with implementing secure data engineering practices for an AI system that handles customer financial information. Which of the following practices should you implement to protect the data during both storage and processing? - Use of public S3 buckets for data storage - Encrypt data at rest and in transit - Data obfuscation techniques - Data replication **CORRECT:** "Encrypt data at rest and in transit" is the correct answer. Encrypting data at rest and in transit is one of the most important practices for securing customer financial information. Encryption at rest ensures that data is protected when stored, while encryption in transit protects data when it is being transferred between systems. This ensures that unauthorized individuals cannot access sensitive information during storage or transmission. AWS services, such as AWS Key Management Service (KMS), can be used to manage encryption keys and apply encryption across various data stores, including Amazon S3 and RDS, as well as during transfers using SSL/TLS. **INCORRECT:** "Data obfuscation techniques" is incorrect. Data obfuscation hides sensitive information, but it is not a comprehensive security solution like encryption for protecting data during storage and transmission. **INCORRECT:** "Data replication" is incorrect. Data replication ensures data availability by copying data across systems, but it does not inherently protect the data from unauthorized access or secure it. **INCORRECT:** "Use of public S3 buckets for data storage" is incorrect. Using public S3 buckets for storing sensitive data is a security risk, as it can expose data to unauthorized users. Private, encrypted storage should be used instead. **References:** https://docs.aws.amazon.com/whitepapers/latest/logical-separation/encrypting-data-at-rest-and--in-transit.html Domain: Security, Compliance, and Governance for AI Solutions --- #### 37. A financial institution is fine-tuning a foundation model to improve fraud detection. They have a large dataset, but most of the transactions are legitimate, making fraudulent transactions rare. They are concerned about the model's performance in detecting these rare cases. What should the institution focus on to improve the fine-tuning process for detecting fraud? - Use continuous pre-training on legitimate transactions - Balance the dataset by oversampling fraudulent transactions - Remove fraudulent data to simplify the model - Curate only legitimate transaction data **CORRECT:** "Balance the dataset by oversampling fraudulent transactions" is the correct answer. To improve the model's ability to detect rare fraudulent transactions, the institution should balance the dataset by oversampling the fraudulent cases or using techniques such as synthetic data generation. This helps the model learn from a more balanced distribution of legitimate and fraudulent transactions, improving its sensitivity to detecting the rare fraud cases. Without balancing, the model may become biased towards the majority class (legitimate transactions) and struggle to identify the minority class (fraud). Balancing the dataset ensures the model is more effective in recognizing patterns related to fraudulent behavior. **INCORRECT:** "Curate only legitimate transaction data" is incorrect. Focusing solely on legitimate transaction data would not help the model improve its detection of rare fraudulent cases. The model needs exposure to fraudulent data to learn the necessary patterns for detection. **INCORRECT:** "Use continuous pre-training on legitimate transactions" is incorrect. Pre-training on legitimate transactions alone would not address the issue of detecting rare fraudulent activities. The model must learn from both legitimate and fraudulent transactions to improve fraud detection. **INCORRECT:** "Remove fraudulent data to simplify the model" is incorrect. Removing fraudulent data would undermine the goal of improving fraud detection. The model needs to be trained on fraudulent data to effectively recognize and detect these transactions. **References:** https://aws.amazon.com/solutions/implementations/fraud-detection-using-machine-learning Domain: Applications of Foundation Models --- #### 38. You are working on a machine learning model that categorizes sensitive legal documents. During evaluation, you notice inconsistencies in the labeled data. What is the best approach to improve label quality and ensure the model's trustworthiness? - Increase the size of the dataset without reviewing label quality. - Use unsupervised learning to address inconsistencies without modifying labels. - Rely solely on automated tools like SageMaker Clarify for bias detection. - Conduct human audits to manually verify and correct any labeling errors. **CORRECT:** "Conduct human audits to manually verify and correct any labeling errors" is the correct answer. The best approach to improve label quality and ensure the model's trustworthiness is to conduct human audits to manually verify and correct any labeling errors. Inconsistent or inaccurate labels can significantly impact the performance and reliability of a machine learning model, especially when dealing with sensitive legal documents. Human auditors, particularly subject matter experts, can ensure that the data is correctly labeled, leading to more accurate model predictions. Manual auditing allows for a detailed review of complex or ambiguous cases, ensuring the model is trained on high-quality, reliable data. This approach helps to avoid model errors, boosts trustworthiness, and maintains compliance with regulatory standards, particularly in fields like law, where precision is critical. **INCORRECT:** "Increase the size of the dataset without reviewing label quality" is incorrect. Increasing the dataset size without addressing labeling inconsistencies will not solve the underlying problem of inaccurate labels. The model would continue to learn from poor-quality data, leading to unreliable predictions. **INCORRECT:** "Rely solely on automated tools like SageMaker Clarify for bias detection" is incorrect. While automated tools like SageMaker Clarify are helpful in detecting bias, they are not designed to correct labeling errors. Human intervention is still required to manually review and correct inconsistencies in labeled data. **INCORRECT:** "Use unsupervised learning to address inconsistencies without modifying labels" is incorrect. Unsupervised learning does not directly address labeling inconsistencies since it works without labeled data. To improve label quality, manual correction of the labels is necessary. **References:** https://aws.amazon.com/blogs/machine-learning/setting-up-human-review-of-your-nlp-based-entity-recognition-models-with-amazon-sagemaker-ground-truth-amazon-comprehend-and-amazon-a2i https://docs.aws.amazon.com/sagemaker/latest/dg/a2i-use-augmented-ai-a2i-human-review-loops.html Domain: Guidelines for Responsible AI --- #### 39. A financial services company is developing AI-powered tools to support fraud detection, automated document processing, and personalized financial advice. The development team is evaluating Amazon Bedrock and Amazon SageMaker JumpStart to speed up model integration and development. They aim to reduce infrastructure overhead while maintaining the ability to experiment with and fine-tune models, especially for tasks like analyzing transaction patterns and summarizing financial documents. To make an informed decision, the team needs to understand how these services differ in terms of access to pre-trained models and their support for customization and deployment. Which of the following best describes this use case? - Amazon SageMaker JumpStart is a serverless solution, whereas Amazon Bedrock requires users to manage their own compute infrastructure. - Amazon SageMaker JumpStart provides access to proprietary foundation models via API, whereas Amazon Bedrock requires users to bring their own models. - Amazon Bedrock provides API-based access to foundation models without managing infrastructure, while SageMaker JumpStart offers pre-built solutions and notebooks for training and fine-tuning. - Both Amazon Bedrock and SageMaker JumpStart allow direct modification of the underlying model architecture. **CORRECT:** "Amazon Bedrock provides API-based access to foundation models without managing infrastructure, while SageMaker JumpStart offers pre-built solutions and notebooks for training and fine-tuning" is the correct answer. Amazon Bedrock is a fully managed service that allows developers to access foundation models (FMs) from providers like Anthropic, AI21 Labs, Meta, and Amazon through a simple API—without managing any underlying infrastructure. It's ideal for those who want quick, serverless access to powerful generative AI models. On the other hand, Amazon SageMaker JumpStart is part of the SageMaker platform and provides pre-built models, example notebooks, and end-to-end solutions that help users quickly start with machine learning. JumpStart supports model training and fine-tuning, enabling customization using your own data. This makes JumpStart suitable for users who need to build and tailor models within a managed but flexible development environment. So, the primary difference is Bedrock focuses on easy API access to hosted foundation models, while JumpStart provides a customizable ML development environment with training capabilities. **INCORRECT:** "Amazon SageMaker JumpStart provides access to proprietary foundation models via API, whereas Amazon Bedrock requires users to bring their own models" is incorrect. Amazon Bedrock provides access to proprietary foundation models via API, not JumpStart. Also, Bedrock users do not need to bring their own models; the service hosts multiple foundation models from different providers. **INCORRECT:** "Both Amazon Bedrock and SageMaker JumpStart allow direct modification of the underlying model architecture" is incorrect. Neither service allows direct editing of the underlying model architecture. Both services support customization (e.g., fine-tuning or prompt engineering), but not architectural changes like modifying neural network layers. **INCORRECT:** "Amazon SageMaker JumpStart is a serverless solution, whereas Amazon Bedrock requires users to manage their own compute infrastructure" is incorrect. This is the opposite of the truth. Amazon Bedrock is serverless, managing infrastructure for model hosting. SageMaker JumpStart, while managed, may require users to configure compute resources for training or deployment. **References:** https://docs.aws.amazon.com/bedrock/latest/userguide/what-is-bedrock.html https://docs.aws.amazon.com/sagemaker/latest/dg/whatis.html Domain: Applications of Foundation Models --- #### 40. A company wants to build an internal generative AI chatbot to provide customer support. The company must comply with strict data residency requirements and wants to avoid sending data to external APIs or third-party services. Which advantage of using AWS Generative AI services addresses this need? - Automatic cost-optimization for large model deployments - Open-source community support and documentation - Built-in transparency dashboards for data lineage - Regional hosting capabilities for data residency **CORRECT:** "Regional hosting capabilities for data residency" is the correct answer. AWS offers regional hosting for generative AI models, ensuring that customer data remains within a specified geographic location to meet data residency and compliance requirements. Services like Amazon Bedrock and Amazon SageMaker allow businesses to train, fine-tune, and deploy AI models within their chosen AWS region, preventing sensitive data from being transferred to external APIs or third-party providers. This is crucial for organizations in regulated industries such as finance, healthcare, and government, where compliance with regulations like GDPR, HIPAA, and local data protection laws is mandatory. **INCORRECT:** "Open-source community support and documentation" is incorrect. While AWS supports open-source AI frameworks like Hugging Face Transformers and PyTorch, this does not directly address data residency concerns. Open-source community support is beneficial for development but does not ensure compliance with strict data protection regulations. **INCORRECT:** "Automatic cost-optimization for large model deployments" is incorrect. AWS provides cost optimization chips like AWS Inferentia, Amazon EC2 Spot Instances, and SageMaker Model Optimization, but cost reduction is unrelated to data residency concerns. Ensuring compliance requires regional model hosting and data governance, which is a separate concern from cost savings. **INCORRECT:** "Built-in transparency dashboards for data lineage" is incorrect. AWS offers some tools for AI explainability and model monitoring, such as Amazon SageMaker Clarify, but these primarily focus on bias detection and model interpretability rather than enforcing data residency requirements. Transparency dashboards do not control where data is stored or processed. **References:** https://aws.amazon.com/compliance/data-privacy https://d1.awsstatic.com/whitepapers/compliance/Data_Residency_Whitepaper.pdf Domain: Fundamentals of Generative AI --- #### 41. As part of your MLOps strategy, you want to ensure that your machine learning models are scalable and can handle increasing data loads in production. Which of the following practices will help you achieve this? - Experimentation with different models - Continuous hyperparameter tuning - Manual deployment processes - Building systems that scale horizontally **CORRECT:** "Building systems that scale horizontally" is the correct answer. Horizontal scalability refers to the ability to add more instances or nodes to distribute the load across multiple machines, ensuring the system can handle a larger volume of data or requests. This approach is crucial in production environments where data loads can grow over time, and the system needs to scale out to maintain performance and reliability. In the context of MLOps, ensuring that your infrastructure can scale horizontally means that your models can serve more predictions, retrain on larger datasets, and process data efficiently as the demand increases. AWS services like Amazon SageMaker support horizontal scaling by allowing the deployment of models across multiple instances. **INCORRECT:** "Experimentation with different models" is incorrect. Experimenting with different models is useful for improving model performance, but it doesn't specifically address the issue of scalability in handling large data loads. **INCORRECT:** "Manual deployment processes" is incorrect. Manual deployment processes are not scalable, as they require human intervention and do not ensure that the system can handle growing data loads efficiently. Automated deployment processes are generally preferred for scalability. **INCORRECT:** "Continuous hyperparameter tuning" is incorrect. Continuous hyperparameter tuning helps optimize model performance but does not directly impact the scalability of the system. Scalability is more about handling increased data loads rather than tuning model parameters. **References:** https://wa.aws.amazon.com/wellarchitected/2020-07-02T19-33-23/wat.concept.horizontal-scaling.en.html Domain: Fundamentals of AI and ML --- #### 42. An automotive company is developing a domain-specific QA system that uses Amazon Bedrock. They want to enable agents to handle multi-step tasks, such as gathering parts inventory data from Amazon DynamoDB, retrieving manufacturing specs from Amazon S3, and then composing a final user-facing response. Which core advantage do Bedrock agents provide in this workflow? - Agents can orchestrate multiple service calls and data lookups during a single conversational flow. - Agents eliminate the need for any fine-tuning or prompt engineering. - Agents disable negative prompts and chain-of-thought reasoning to reduce complexity. - Agents automatically create new foundation models for each type of query. **CORRECT:** "Agents can orchestrate multiple service calls and data lookups during a single conversational flow" is the correct answer. Amazon Bedrock agents enable automation of multi-step tasks by orchestrating API calls, retrieving data from different sources (like Amazon DynamoDB and Amazon S3), and generating coherent responses. These agents enhance the workflow by dynamically handling data retrieval and reasoning, ensuring that the right information is fetched and processed efficiently. Instead of requiring users to manually query different services, Bedrock agents streamline interactions by managing the entire process within a single conversational flow. This reduces development complexity and improves response accuracy while keeping the system cost-effective and scalable. **INCORRECT:** "Agents eliminate the need for any fine-tuning or prompt engineering" is incorrect. While Bedrock agents reduce the need for fine-tuning in many cases, prompt engineering is still useful for optimizing responses. Agents rely on well-structured prompts and logic to execute tasks effectively. **INCORRECT:** "Agents automatically create new foundation models for each type of query" is incorrect. Bedrock agents do not create new models. Instead, they use existing foundation models and enhance their functionality by integrating API calls and external data retrieval. **INCORRECT:** "Agents disable negative prompts and chain-of-thought reasoning to reduce complexity" is incorrect. Bedrock agents do not disable chain-of-thought reasoning. In fact, they often leverage structured reasoning to execute multi-step workflows effectively. **References:** https://docs.aws.amazon.com/bedrock/latest/userguide/agents.html Domain: Applications of Foundation Models --- #### 43. You are working with a team that needs to experiment with and fine-tune foundation models to generate personalized content for users in a cost-efficient environment. Which AWS service provides an interactive playground for testing generative AI models? - Amazon SageMaker JumpStart - Amazon Q - Amazon SageMaker Data Wrangler - PartyRock, an Amazon Bedrock Playground **CORRECT:** "PartyRock, an Amazon Bedrock Playground" is the correct answer. PartyRock, part of the Amazon Bedrock Playground, is a service that allows teams to experiment with and fine-tune foundation models in an interactive environment. This service is designed to let users test generative AI models easily, making it ideal for teams who want to generate personalized content. With PartyRock, you can experiment with different models without worrying about managing infrastructure, ensuring a cost-efficient way to develop and test your AI solutions. **INCORRECT:** "Amazon Q" is incorrect. Amazon Q is a generative AI–powered assistant designed to help both developers and business users. It assists developers by generating code, testing, debugging, and improving workflows with advanced planning and reasoning. It does not provide an interactive playground for testing generative AI models. **INCORRECT:** "Amazon SageMaker Data Wrangler" is incorrect. Amazon SageMaker Data Wrangler is a tool for preparing and transforming data for machine learning. It does not provide an interactive playground for testing generative AI models. **INCORRECT:** "Amazon SageMaker JumpStart" is incorrect. Amazon SageMaker JumpStart provides access to pre-built machine learning models and solutions but does not offer an interactive playground environment specifically designed for testing and fine-tuning generative AI models. **References:** https://partyrock.aws Domain: Fundamentals of Generative AI --- #### 44. A retail company wants to automate customer support using generative AI. The support system should be able to interact with customers, understand their queries, retrieve order details from an external database, and respond with accurate information, all without human intervention. The company is exploring Amazon Bedrock Agents to implement this intelligent automation. Which statement correctly describes the behavior of an agent in Amazon Bedrock? - Agents embed real-time data into foundation models during pretraining. - Agents are used to translate prompts into embeddings for faster inference. - Agents apply supervised learning to optimize foundation model weights during inference. - Agents orchestrate interactions between the user, the foundation model, and external APIs to accomplish tasks. **CORRECT:** "Agents orchestrate interactions between the user, the foundation model, and external APIs to accomplish tasks" is the correct answer. Amazon Bedrock Agents are designed to help applications go beyond simple text generation by allowing them to interact intelligently with external systems and APIs. An agent uses a foundation model (FM) under the hood to understand user input and respond in a conversational manner. What makes it powerful is that it can be configured with instructions and actions, which enable the model to call external APIs (such as a customer order database or inventory system) during the conversation. This orchestration allows the agent to retrieve relevant information and provide accurate, dynamic responses based on real data. For example, if a user asks, "Where is my order?", the agent can fetch order status by calling an API and then use the FM to deliver a natural-language reply. **INCORRECT:** "Agents are used to translate prompts into embeddings for faster inference" is incorrect. This describes embedding models or vector-based search systems, not Bedrock Agents. Embeddings are numerical representations of text used for similarity search or semantic matching. Agents focus on completing tasks, not generating embeddings. **INCORRECT:** "Agents apply supervised learning to optimize foundation model weights during inference" is incorrect. Agents do not perform training or fine-tuning at inference time. They use pre-trained foundation models and predefined logic (like instructions and API calls) to interact and generate responses. Supervised learning is done during model training, not during agent operation. **INCORRECT:** "Agents embed real-time data into foundation models during pretraining" is incorrect. Pretraining happens before deployment, using large datasets. Real-time data interaction (like querying an API) happens during inference or live usage—not during pretraining. Agents work with pre-trained models and don't modify them on the fly. **References:** https://docs.aws.amazon.com/bedrock/latest/userguide/agents.html Domain: Applications of Foundation Models --- #### 45. A retail company needs to quickly develop a model that predicts customer churn based on historical data. They are looking for a solution that automates the entire process from data preprocessing to model deployment. Which AWS feature should they use to meet this requirement? - Amazon SageMaker Autopilot - Amazon SageMaker Model Monitor - Amazon SageMaker Feature Store - Amazon SageMaker Ground Truth **CORRECT:** "Amazon SageMaker Autopilot" is the correct answer. Amazon SageMaker Autopilot is a tool that automates the process of building, training, and tuning machine learning models, while still providing full visibility and control over each step. It automatically explores different data preprocessing techniques, algorithms, and hyperparameters to find the best performing model for a given dataset. Unlike other AutoML tools, SageMaker Autopilot allows users to review and modify the generated models, offering both automation and transparency. This enables users, including those without deep machine learning expertise, to create highly accurate models quickly and efficiently while retaining the flexibility to customize as needed. **INCORRECT:** "Amazon SageMaker Ground Truth" is incorrect. SageMaker Ground Truth is primarily used for creating labeled datasets through data annotation, not for automating the full machine learning lifecycle. **INCORRECT:** "Amazon SageMaker Model Monitor" is incorrect. SageMaker Model Monitor tracks and monitors the performance of models in production to detect issues like data drift. It does not handle the automated development of models from data preprocessing to deployment. **INCORRECT:** "Amazon SageMaker Feature Store" is incorrect. SageMaker Feature Store is used to store and manage machine learning features but does not automate the entire process from preprocessing to deployment. **References:** https://docs.aws.amazon.com/sagemaker/latest/dg/autopilot-automate-model-development.html Domain: Fundamentals of AI and ML --- #### 46. A company needs to ensure that its AI system's data moves through various stages such as collection, processing, storage, and deletion, in a controlled and compliant manner. Which data governance strategy does this process describe? - Data Lifecycle Management - Data Retention - Data Logging - Data Monitoring **CORRECT:** "Data Lifecycle Management" is the correct answer. Data lifecycle management refers to the process of controlling how data moves through its various stages, including collection, processing, storage, and eventual deletion. This governance strategy ensures that data is handled in a compliant and structured manner throughout its entire lifespan. By implementing data lifecycle management, companies can ensure that sensitive data is managed according to regulatory requirements, from creation to disposal, while maintaining data integrity and security. This approach is essential for AI systems handling vast amounts of data, helping to automate policies and reduce risks. **INCORRECT:** "Data Monitoring" is incorrect. Data monitoring focuses on observing data streams for anomalies and irregularities in real-time. While important for system integrity, it does not encompass the entire lifecycle of data. **INCORRECT:** "Data Retention" is incorrect. Data retention governs how long data is stored before it is archived or deleted. It is a component of the data lifecycle but does not describe the full process. **INCORRECT:** "Data Logging" is incorrect. Data logging involves recording actions taken on data, such as access, modifications, and deletions. While useful for traceability, it does not manage the full lifecycle of data. **References:** https://docs.aws.amazon.com/prescriptive-guidance/latest/modern-data-centric-use-cases/data-lifecycle.html https://docs.aws.amazon.com/prescriptive-guidance/latest/oracle-exadata-blueprint/discovery-lifecycle.html Domain: Security, Compliance, and Governance for AI Solutions --- #### 47. An insurance company deployed a model to assess customer risk profiles. Over time, the model's predictions became inconsistent due to new claim trends and customer behavior. To restore model performance, engineers implemented a pipeline that automatically incorporates new labeled data to refine predictions. Which process does this describe in an MLOps environment? - Model checkpointing to preserve inference reproducibility - Model retraining to ensure continued alignment with data patterns - Real-time model tuning based on user feedback - Enhanced inference acceleration using batch prediction techniques **CORRECT:** "Model retraining to ensure continued alignment with data patterns" is the correct answer. Model retraining is the process of updating an existing machine learning model by using new, often labeled, data to adapt it to recent trends or changes in real-world behavior. In an MLOps environment, this process is critical when the data patterns shift over time—also known as data drift or concept drift. In the case of the insurance company, customer claim trends and behaviors have changed, leading to inaccurate predictions. By implementing a pipeline that regularly adds new labeled data and retrains the model, the team ensures that the model stays aligned with the current data distribution and remains accurate. This is a common practice to maintain the reliability and effectiveness of AI systems over time. **INCORRECT:** "Enhanced inference acceleration using batch prediction techniques" is incorrect. Batch prediction techniques are used to improve the speed and efficiency of processing large amounts of input data in groups. While helpful for reducing computational costs, batch prediction does not involve updating or improving the model based on new data, so it doesn't solve the issue of prediction inconsistency. **INCORRECT:** "Model checkpointing to preserve inference reproducibility" is incorrect. Model checkpointing is a strategy used during training to save a snapshot of the model's state. It allows you to resume training or roll back to a previous version if needed. However, it doesn't involve learning from new data or restoring model performance due to changing trends. **INCORRECT:** "Real-time model tuning based on user feedback" is incorrect. Real-time model tuning implies adjusting model parameters live during prediction based on immediate feedback. While interactive and adaptive, this is more common in reinforcement learning scenarios or personalization engines, and it's not the standard approach for restoring accuracy in predictive models that deal with structured data like insurance claims. **References:** https://docs.aws.amazon.com/machine-learning/latest/dg/retraining-models-on-new-data.html Domain: Fundamentals of AI and ML --- #### 48. An enterprise software firm is planning to adopt Amazon Q Developer to streamline internal development processes and boost engineering productivity. The IT leadership wants to understand which features will offer the most impact. Which of the following represents the capabilities of Amazon Q Developer? (Select TWO) - Generate AI model weights for training deep learning systems. - Replace DevOps engineers with autonomous agents. - Suggest relevant code and help fix common programming errors. - Replace AWS CLI for all infrastructure tasks. - Automate repetitive development tasks such as testing and documentation. **CORRECT:** "Suggest relevant code and helps fix common programming errors" is a correct answer. Amazon Q Developer is an AI-powered assistant designed to help developers write better code faster. It integrates with IDEs (like VS Code) and can suggest code completions, detect and help fix common programming errors, and even explain code snippets. This reduces debugging time and helps developers stay focused on core logic rather than syntax issues or repetitive fixes. **CORRECT:** "Automate repetitive development tasks such as testing and documentation" is also a correct answer. One of Amazon Q Developer's strengths is automating routine tasks. It can generate unit tests, document code automatically, and even assist in writing scripts or configuration files. These time-saving features allow developers to focus on building features rather than managing boilerplate code or writing repetitive documentation. **INCORRECT:** "Replace DevOps engineers with autonomous agents" is incorrect. Amazon Q Developer enhances productivity but does not replace DevOps engineers. It assists with coding and development workflows, but human expertise is still essential for deployment, infrastructure, and CI/CD pipelines. **INCORRECT:** "Replace AWS CLI for all infrastructure tasks" is incorrect. Amazon Q Developer can help write and explain commands or infrastructure-as-code scripts but does not replace the AWS CLI. It complements existing tools rather than replacing them entirely. **INCORRECT:** "Generates AI model weights for training deep learning systems" is incorrect. Amazon Q Developer does not train deep learning models or generate model weights. Its primary focus is assisting with software development and productivity, not machine learning model training. **References:** https://aws.amazon.com/q/developer https://docs.aws.amazon.com/amazonq/latest/qdeveloper-ug/what-is.html Domain: Fundamentals of Generative AI --- #### 49. A retail company plans to develop a system that can automatically forecast product demand. The team decides to apply machine learning techniques to achieve this goal. Which of the following statements describe the primary role of machine learning in this context? (Select TWO.) - Machine learning eliminates the need for any human supervision or validation. - Machine learning replicates human brain functionality exactly to simulate human-level decision-making. - Machine learning is used to develop Large Language Models for text-based applications. - Machine learning enables the system to learn from historical data to identify patterns and make future predictions. - Machine learning supports both supervised and unsupervised learning to solve various business problems. **CORRECT:** "Machine learning enables the system to learn from historical data to identify patterns and make future predictions" is a correct answer. Machine learning is a branch of artificial intelligence (AI) that allows computer systems to automatically learn from historical data without being explicitly programmed with fixed rules. Instead of relying on manual logic, machine learning models analyze large amounts of data to detect patterns and relationships. These learned patterns enable the system to make accurate predictions or decisions about new, unseen data. In the case of demand forecasting, machine learning uses past sales, seasonality, and customer behavior data to predict future product demand, helping businesses make better inventory and supply chain decisions. **CORRECT:** "Machine learning supports both supervised and unsupervised learning to solve various business problems" is also a correct answer. Machine learning offers different types of learning approaches. Supervised learning is used when the model is trained on labeled data to predict future outcomes, such as forecasting product demand. Unsupervised learning is used to find hidden patterns in unlabeled data, such as grouping customers based on buying behavior. Both methods are widely applied across industries to solve different types of business challenges, from customer segmentation to fraud detection. This flexibility makes ML valuable for a wide range of applications. **INCORRECT:** "Machine learning replicates human brain functionality exactly to simulate human-level decision-making" is incorrect. While machine learning is inspired by how humans learn from experience, it does not replicate human brain functionality exactly. ML models use algorithms and mathematical functions to process data and make predictions. They are excellent at pattern recognition and automation but do not possess human intuition, reasoning, or consciousness. Claiming that ML simulates human-level decision-making exactly is misleading. **INCORRECT:** "Machine learning eliminates the need for any human supervision or validation" is incorrect. Machine learning systems require human involvement at various stages, such as defining objectives, preparing data, training models, and validating results. Even after deployment, human oversight is needed to monitor performance and make adjustments as needed. Machine learning can automate tasks, but it does not fully eliminate the need for human supervision. This statement is misleading and incorrect. **INCORRECT:** "Machine learning is used to develop Large Language Models for text-based applications" is incorrect. Large Language Models (LLMs) are a specific application of machine learning, primarily used in natural language processing (NLP). While this is an important area, it does not describe the primary role of ML in the scenario of product demand forecasting, which involves analyzing structured data like sales records and inventory levels, not text-based data. **References:** https://aws.amazon.com/what-is/machine-learning https://aws.amazon.com/ai/machine-learning Domain: Fundamentals of AI and ML --- #### 50. A retail company wants to use generative AI to automate customer service chat responses. They are concerned about providing accurate answers to customer queries, particularly for product recommendations. What would be the main limitation they should consider when choosing a generative AI model for this task? - Generative AI models are incapable of understanding natural language. - Generative AI models are limited to text-based outputs and cannot recommend products. - Generative AI models always provide the same response to the same input. - Generative AI models may hallucinate and provide inaccurate responses. **CORRECT:** "Generative AI models may hallucinate and provide inaccurate responses" is the correct answer. Generative AI models, such as large language models (LLMs), can sometimes produce responses that are plausible-sounding but factually incorrect or nonsensical—a phenomenon known as "hallucination." In the context of customer service and product recommendations, hallucinations can lead to misinformation, customer dissatisfaction, and potential harm to the company's reputation. Therefore, the main limitation the retail company should consider is the risk of these models providing inaccurate or misleading information to customers. Mitigating this issue may involve techniques like fine-tuning the model with domain-specific data, implementing retrieval-augmented generation, or incorporating human oversight. **INCORRECT:** "Generative AI models always provide the same response to the same input" is incorrect. Generative AI models often produce varying responses to the same input due to their probabilistic nature. They use randomness in their decoding processes (e.g., through sampling methods like top-k or nucleus sampling), which allows them to generate diverse and creative outputs. While this variability can be adjusted using parameters like "temperature," it is not accurate to say they always provide the same response. **INCORRECT:** "Generative AI models are incapable of understanding natural language" is incorrect. Generative AI models are specifically designed to process and generate natural language. They are trained on vast amounts of textual data to understand language patterns, grammar, and context. While they don't "understand" language in a human sense, they are highly capable of interpreting natural language inputs and producing coherent and contextually relevant outputs. **INCORRECT:** "Generative AI models are limited to text-based outputs and cannot recommend products" is incorrect. While generative AI models primarily produce text-based outputs, they can indeed recommend products by generating text that includes product suggestions. By integrating the AI model with a product database or knowledge base, it can provide personalized recommendations based on customer queries. Therefore, it's incorrect to say they cannot recommend products. **References:** https://aws.amazon.com/blogs/publicsector/generative-ai-understand-the-challenges-to-realize-the-opportunities Domain: Fundamentals of Generative AI --- #### 51. A healthcare provider wants to develop an AI-powered system that summarizes lengthy medical reports into concise summaries for doctors. The system should extract relevant information while maintaining medical accuracy. **Question** The AI-generated medical summaries should not alter or misrepresent patient conditions. What best practice can help maintain medical accuracy and prevent misleading information? - Fine-tuning the AI model using reinforcement learning with human feedback (RLHF) - Using a general-purpose summarization model without domain-specific training - Reducing the training dataset size to limit model complexity - Allowing the model to generate unrestricted summaries without human review **CORRECT:** "Fine-tuning the AI model using reinforcement learning with human feedback (RLHF)" is the correct answer. Fine-tuning with Reinforcement Learning from Human Feedback (RLHF) ensures that AI-generated medical summaries maintain accuracy and do not misrepresent patient conditions. In this approach, medical experts review model outputs and provide feedback, helping the AI refine its responses. This iterative process minimizes hallucinations, enhances medical reliability, and aligns summaries with real-world healthcare standards. RLHF is crucial for ensuring that AI-generated summaries uphold ethical and legal healthcare guidelines. **INCORRECT:** "Allowing the model to generate unrestricted summaries without human review" is incorrect. Letting the model generate summaries without oversight increases the risk of hallucinations, misinformation, or omission of key medical details. Human review is essential to ensure accuracy and maintain trust in AI-assisted healthcare applications. **INCORRECT:** "Using a general-purpose summarization model without domain-specific training" is incorrect. General-purpose summarization models lack medical expertise and may not understand complex terminology or critical relationships in medical reports. Without domain-specific training, the summaries could be incomplete or misleading. **INCORRECT:** "Reducing the training dataset size to limit model complexity" is incorrect. Reducing the dataset size can negatively impact the model's learning, leading to poor generalization and lower accuracy. Instead of limiting data, high-quality and domain-specific medical datasets should be used to train a reliable summarization model. **References:** https://aws.amazon.com/what-is/reinforcement-learning-from-human-feedback Domain: Guidelines for Responsible AI --- #### 52. A research lab is building a machine learning model for climate prediction and wants to ensure the model is transparent and can be easily understood by other researchers. They are considering using an open-source model. What is a key benefit of using an open-source model for transparency? - Open-source models provide built-in explanations for all predictions. - Open-source models generally have higher accuracy than proprietary models. - Open-source models allow users to inspect and modify the code, making the model more transparent. - Open-source models are automatically licensed for commercial use without restrictions. **CORRECT:** "Open-source models allow users to inspect and modify the code, making the model more transparent" is the correct answer. A key benefit of using open-source models for transparency is that they allow users to inspect and modify the code. This makes it easier for researchers and developers to understand how the model works, evaluate its decision-making process, and ensure that it aligns with their goals or methodologies. Open-source models encourage collaboration and sharing, enabling other researchers to verify and improve the model, which is essential in scientific research like climate prediction. **INCORRECT:** "Open-source models generally have higher accuracy than proprietary models" is incorrect. Open-source models do not necessarily have higher accuracy than proprietary models. Accuracy depends on the model's design, data, and training, not whether it is open-source. **INCORRECT:** "Open-source models are automatically licensed for commercial use without restrictions" is incorrect. Open-source models are often subject to specific licenses that may impose restrictions on commercial use. Users should always check the licensing terms before using an open-source model for commercial purposes. **INCORRECT:** "Open-source models provide built-in explanations for all predictions" is incorrect. While open-source models offer transparency through accessible code, they do not automatically provide built-in explanations for every prediction. Explanation features depend on the specific implementation. **References:** https://aws.amazon.com/opensource Domain: Guidelines for Responsible AI --- #### 53. A retail parking facility wants to automatically record the license plate numbers of incoming vehicles and identify if any vehicle from a predefined blacklist enters the premises. Which AWS machine learning service(s) would be most suitable for implementing this solution? - Use Amazon Comprehend to extract vehicle and license plate data from video frames and classify blacklisted vehicles. - Use Amazon Rekognition to detect and extract license plate numbers from real-time video streams, and compare them against a blacklist stored in a database. - Use Amazon Kendra to search for license plate numbers in indexed documents and match them with the blacklist. - Use Amazon Bedrock to generate responses that determine whether a vehicle is blacklisted based on its license plate. **CORRECT:** "Use Amazon Rekognition to detect and extract license plate numbers from real-time video streams, and compare them against a blacklist stored in a database" is the correct answer. Amazon Rekognition is a powerful AWS machine learning service that enables image and video analysis. It can detect objects, scenes, and text in images and videos, which includes recognizing and extracting license plate numbers from video frames. For a retail parking facility, Amazon Rekognition can be used to process real-time video streams of incoming vehicles. It will identify and extract the text from the license plates and send this data to a backend system. Once the license plate number is extracted, it can be compared against a predefined blacklist stored in a database such as Amazon DynamoDB or Amazon RDS. If a match is found, an alert or automatic action can be triggered. This approach is scalable, accurate, and aligns with AWS best practices for computer vision-based applications. **INCORRECT:** "Use Amazon Comprehend to extract vehicle and license plate data from video frames and classify blacklisted vehicles" is incorrect. Amazon Comprehend is a Natural Language Processing (NLP) service used to analyze and extract insights from textual data such as documents or sentences. It cannot process video or image data, and therefore it's not capable of extracting license plate numbers from video frames. Using Comprehend for visual tasks like this would not be appropriate. **INCORRECT:** "Use Amazon Kendra to search for license plate numbers in indexed documents and match them with the blacklist" is incorrect. Amazon Kendra is an intelligent search service that enables users to search unstructured data across multiple content repositories. While it is useful for searching through text documents and knowledge bases, it is not designed for real-time image or video analysis, and it does not support object detection or text extraction from media files. **INCORRECT:** "Use Amazon Bedrock to generate responses that determine whether a vehicle is blacklisted based on its license plate" is incorrect. Amazon Bedrock allows you to build and scale generative AI applications using foundation models. It is not designed to analyze video streams or detect objects like license plates. While it can assist in generating text-based insights or summaries, it cannot directly process video or extract license plate data from images. **References:** https://docs.aws.amazon.com/rekognition/latest/dg/what-is.html https://docs.aws.amazon.com/rekognition/latest/dg/text-detection.html Domain: Applications of Foundation Models --- #### 54. Which of the following best describes a neural network in the context of machine learning? - A system that learns through trial and error using rewards and penalties. - A computational model inspired by the structure of the human brain, composed of layers of interconnected nodes. - A technique that automatically adjusts hyperparameters during model training. - A method of classifying data into predefined categories. **CORRECT:** "A computational model inspired by the structure of the human brain, composed of layers of interconnected nodes" is the correct answer. A neural network in machine learning is a computational model that mimics the structure of the human brain. It consists of layers of nodes (also called neurons) that are interconnected. These layers typically include an input layer, one or more hidden layers, and an output layer. Neural networks are used to detect patterns in data, and they are the foundation for more complex deep learning models. Each connection in the network has a weight that adjusts during training to minimize errors, allowing the network to learn from data. **INCORRECT:** "A system that learns through trial and error using rewards and penalties" is incorrect. This describes reinforcement learning, where an agent learns by interacting with an environment and receiving feedback in the form of rewards and penalties, but it is not specific to neural networks. **INCORRECT:** "A method of classifying data into predefined categories" is incorrect. This describes classification, which can be performed by various algorithms, including neural networks, but this option does not specifically define what a neural network is. **INCORRECT:** "A technique that automatically adjusts hyperparameters during model training" is incorrect. This describes hyperparameter tuning, which is a process for optimizing model parameters but is not a defining characteristic of neural networks. **References:** https://aws.amazon.com/what-is/neural-network https://docs.aws.amazon.com/machine-learning/latest/dg/machine-learning-concepts.html Domain: Fundamentals of AI and ML --- #### 55. An e-commerce business is considering generative AI to automate its product description generation. They are looking for a solution that can streamline the process with minimal technical setup and user interaction. What is one of the main advantages of using generative AI for this purpose? - Generative AI only functions well with structured and predefined data. - Generative AI cannot automate product description tasks and requires manual writing. - Generative AI is highly complex to set up and requires constant manual input. - Generative AI simplifies content creation by automating the process with little technical expertise needed. **CORRECT:** "Generative AI simplifies content creation by automating the process with little technical expertise needed" is the correct answer. Generative AI models are designed to simplify and speed up content creation. These models can process minimal input, such as product details or key features, and automatically generate well-structured, SEO-friendly descriptions. For e-commerce businesses, this significantly reduces the time and cost of manual writing and doesn't require deep technical expertise for setup. Many AI solutions are designed with user-friendly interfaces, making them accessible to non-technical users. This makes generative AI highly effective for streamlining content creation tasks. **INCORRECT:** "Generative AI is highly complex to set up and requires constant manual input" is incorrect. Modern generative AI platforms are built to be user-friendly, requiring minimal setup or manual input once trained models are in place. Users can generate content with ease. **INCORRECT:** "Generative AI cannot automate product description tasks and requires manual writing" is incorrect. One of the key advantages of generative AI is its ability to automate tasks like product description generation, reducing the need for manual writing. **INCORRECT:** "Generative AI only functions well with structured and predefined data" is incorrect. Generative AI can work with a variety of data types, including unstructured data, and is capable of learning patterns from large datasets to generate content dynamically. **References:** https://d1.awsstatic.com/psc-digital/2024/gc-600/genai-adv-tdm/The-generative-AI-advantage.pdf Domain: Fundamentals of Generative AI --- #### 56. A media company uses a generative AI model via Amazon Bedrock to create content for social platforms. To ensure that the AI does not produce content related to prohibited or sensitive topics such as violence, hate speech, or misinformation, the team is exploring the use of Guardrails for Amazon Bedrock. Which statement best describes the role of Guardrails in this scenario? - Guardrails block user access to models that are not pre-approved by administrators. - Guardrails allow the model to generate content more quickly by optimizing input formatting. - Guardrails integrate external APIs to improve factual accuracy in generated responses. - Guardrails help prevent the model from generating content related to restricted or sensitive topics. **CORRECT:** "Guardrails help prevent the model from generating content related to restricted or sensitive topics" is the correct answer. Guardrails for Amazon Bedrock are designed to enhance the safety and responsibility of generative AI applications by allowing developers to configure rules that block inappropriate or undesired content. In this case, the media company can use Guardrails to prevent the model from generating outputs that include violence, hate speech, profanity, or misinformation. Guardrails work by scanning both the input and the output of the model, and can stop unsafe content before it's delivered to the user. This is especially important in industries like media and publishing, where reputation and compliance with platform guidelines are critical. By enabling Guardrails, companies can reduce the risk of publishing harmful or non-compliant content. **INCORRECT:** "Guardrails allow the model to generate content more quickly by optimizing input formatting" is incorrect. Guardrails are not focused on performance or speed improvements. They are designed for content safety, not optimization of input or output generation speed. **INCORRECT:** "Guardrails block user access to models that are not pre-approved by administrators" is incorrect. This describes access control or permissions management, not the function of Guardrails. Guardrails do not manage model-level access; they filter the content generated by the model. **INCORRECT:** "Guardrails integrate external APIs to improve factual accuracy in generated responses" is incorrect. Guardrails do not enhance factual accuracy. Instead, they filter unsafe content. Improving factual accuracy would require methods like Retrieval-Augmented Generation (RAG) or fact-checking tools. **References:** https://docs.aws.amazon.com/bedrock/latest/userguide/guardrails.html Domain: Applications of Foundation Models --- #### 57. A robotics company wants to enhance the behavior of its warehouse robots that operate in dynamic environments. They plan to use Reinforcement Learning (RL) to train the robots for better decision-making and adaptability. Which of the following statements are TRUE about Reinforcement Learning? (Select TWO.) - Reinforcement Learning is best suited for analyzing customer sentiments in reviews and chat applications. - Reinforcement Learning is primarily used to generate content such as text, images, or audio. - Reinforcement Learning is widely used in domains such as robotics, game playing, and autonomous vehicles. - Reinforcement Learning enables agents to learn optimal behaviors through trial and error interactions with the environment. - Reinforcement Learning is mainly designed for static data problems like classification and regression. **CORRECT:** "Reinforcement Learning enables agents to learn optimal behaviors through trial and error interactions with the environment" is a correct answer. Reinforcement Learning (RL) is a type of machine learning where an agent or software learns to make decisions by interacting with an environment. It does this through a trial-and-error process, receiving rewards for good actions and penalties for bad ones. Over time, the agent learns the best strategy or policy to achieve its goals. This approach is perfect for situations where there isn't a clear answer, and the agent must figure out what works best through experience. That's why RL is so useful in robotics, where conditions often change and adaptability is key. **CORRECT:** "Reinforcement Learning is widely used in domains such as robotics, game playing, and autonomous vehicles" is also a correct answer. Reinforcement Learning shines in environments where decisions need to be made step-by-step over time. It has been used successfully in robotics (to help robots move and adapt), in games (like training AI agents to beat human players), and in autonomous vehicles (to help cars navigate complex roads). These use cases involve dynamic environments, which are perfect for RL's decision-making and learning approach. **INCORRECT:** "Reinforcement Learning is mainly designed for static data problems like classification and regression" is incorrect. This describes supervised learning, not Reinforcement Learning. Supervised learning deals with fixed datasets and is used for tasks like predicting a label or value. In contrast, RL involves an interactive process with dynamic environments and learning from consequences of actions. **INCORRECT:** "Reinforcement Learning is primarily used to generate content such as text, images, or audio" is incorrect. This is a key feature of generative models like GANs (for images) or large language models (for text). While RL can be combined with these models (e.g., for fine-tuning), content generation is not its primary function. **INCORRECT:** "Reinforcement Learning is best suited for analyzing customer sentiments in reviews and chat applications" is incorrect. Sentiment analysis is typically done using natural language processing (NLP) techniques, often with supervised learning. RL isn't commonly used for this task because there's no interactive environment or reward-based feedback involved in understanding sentiments. **References:** https://aws.amazon.com/what-is/reinforcement-learning Domain: Fundamentals of AI and ML --- #### 58. A hospital research team wants to build a machine learning model to predict patient readmission rates using historical health records. However, the team lacks data science or programming skills. They want a tool that allows them to build and evaluate a model using a graphical interface without writing any code. Which AWS service best fits this requirement? - Amazon Comprehend, which uses machine learning to uncover insights and relationships in ML models. - Amazon SageMaker Canvas, which provides a visual interface for building ML models without coding. - AWS Glue Studio, which offers a visual interface for creating, running, and monitoring ML models without coding. - Amazon Bedrock, which allows developers to build generative AI applications using foundation models without coding. **CORRECT:** "Amazon SageMaker Canvas, which provides a visual interface for building ML models without coding" is the correct answer. Amazon SageMaker Canvas is a machine learning (ML) service designed specifically for users without any coding or data science experience. It provides a no-code, visual interface that allows users to prepare data, build predictive models, and generate insights by simply pointing and clicking. For a hospital research team with limited technical expertise, SageMaker Canvas is an ideal solution because it abstracts away the complexities of machine learning workflows. It supports automatic model creation and evaluation while allowing users to experiment with datasets interactively. This tool enables domain experts like healthcare professionals to make data-driven decisions without needing to learn programming or ML concepts. **INCORRECT:** "Amazon Comprehend, which uses machine learning to uncover insights and relationships in ML models" is incorrect. Amazon Comprehend is a natural language processing (NLP) service designed to extract insights from text data, such as detecting sentiment, identifying key phrases, and recognizing entities. It is not a general-purpose ML model builder and does not offer a visual interface for creating predictive models from structured datasets like patient records. Therefore, it is not suitable for building custom ML models for predicting patient readmissions. **INCORRECT:** "Amazon Bedrock, which allows developers to build generative AI applications using foundation models without coding" is incorrect. Amazon Bedrock enables developers to use pre-trained foundation models from providers like Anthropic, AI21 Labs, and Amazon itself to build generative AI applications. While it simplifies working with generative AI, it still requires some understanding of prompt design and coding. Moreover, Bedrock is intended for use cases involving text generation, summarization, and chatbots, rather than structured data prediction like patient readmission modeling. **INCORRECT:** "AWS Glue Studio, which offers a visual interface for creating, running, and monitoring ML models without coding" is incorrect. AWS Glue Studio is primarily used for building and managing data integration workflows through ETL (Extract, Transform, Load) processes. It's useful for preparing data for analytics or machine learning but does not provide tools to build, train, or evaluate machine learning models. Therefore, it is not suitable for model creation in this context. **References:** https://aws.amazon.com/sagemaker-ai/canvas Domain: Fundamentals of Generative AI --- #### 59. Select and order the following security measures for generative AI applications from LOWEST to HIGHEST level of security enforcement. Each measure should be selected one time. (Select and order THREE.) Note: Select only the correct options, as the type of "Ordering" question is not supported here. - API key-based access control - Attribute-based access control (ABAC) with granular policies - Role-based access control (RBAC) **CORRECT:** The correct order from lowest to highest security level is: 1. API key-based access control 2. Role-based access control (RBAC) 3. Attribute-based access control (ABAC) with granular policies When securing AI applications, different access control mechanisms provide varying levels of security enforcement. Here's the correct order from lowest to highest security level: API Key-Based Access Control – This is the simplest form of security, where access is granted through an API key. While easy to implement, API keys lack granularity and can be misused if exposed. They generally provide all-or-nothing access and do not differentiate between users or roles. Role-Based Access Control (RBAC) – RBAC enhances security by assigning users specific roles with predefined permissions. It ensures that only authorized individuals (e.g., admins, developers, or analysts) can access certain resources. While more secure than API keys, RBAC lacks flexibility for complex, dynamic policies. Attribute-Based Access Control (ABAC) with Granular Policies – ABAC provides the highest level of security by enforcing access rules based on multiple attributes, such as user identity, resource type, and request context. ABAC allows fine-grained control, making it ideal for highly secure AI applications. Choosing the right security measure ensures better access control, data protection, and compliance for AI-powered systems. Domain: Security, Compliance, and Governance for AI Solutions --- #### 60. A music streaming service wants to improve its recommendation engine by tagging songs with multiple mood labels like energetic, sad, or relaxing, based on the song's lyrics and beat. Sometimes a song might fit more than one mood — for example, a song could be both romantic and relaxing. The team needs a machine learning method that can handle this requirement, ensuring songs can have multiple mood tags instead of being forced into just one category. Which ML classification method should the team implement? - Clustering - Multi-class classification - Multi-label classification - Regression **CORRECT:** "Multi-label classification" is the correct answer. Multi-label classification is a type of machine learning where each instance (in this case, a song) can be assigned multiple labels at the same time. For example, a song could be tagged as both "romantic" and "relaxing." This fits the music streaming service's need perfectly because moods can overlap, and a song isn't restricted to just one category. Multi-label classification helps the recommendation engine provide richer and more accurate song suggestions based on multiple moods rather than forcing a single choice. **INCORRECT:** "Multi-class classification" is incorrect. Multi-class classification is used when each instance must be assigned exactly one label from many possible classes. It does not allow multiple labels. Since songs in this case can have more than one mood tag, multi-class classification would not meet the requirement. **INCORRECT:** "Clustering" is incorrect. Clustering is an unsupervised learning method that groups similar items without using predefined labels. While clustering might group songs with similar beats or styles together, it won't assign clear, meaningful mood tags based on existing labels like "romantic" or "relaxing." This method is not suitable for tagging songs with known moods. **INCORRECT:** "Regression" is incorrect. Regression is used to predict continuous numerical values, such as the number of song plays or a popularity score. It is not designed for assigning mood labels, especially when multiple labels per song are needed. **References:** https://docs.aws.amazon.com/machine-learning/latest/dg/types-of-ml-models.html https://aws.amazon.com/blogs/machine-learning/amazon-comprehend-now-supports-multi-label-custom-classification Domain: Fundamentals of AI and ML --- #### 61. An AI system designed to predict employee retention shows a significantly different accuracy for predictions based on department. Some departments have much lower prediction accuracy compared to others. Which technique would best help to identify and address this issue? - Using Amazon Transcribe to improve the training data quality. - Subgroup analysis to evaluate model performance across different departments. - Automated hyperparameter tuning to optimize model accuracy. - Real-time model monitoring with Amazon Polly. **CORRECT:** "Subgroup analysis to evaluate model performance across different departments" is the correct answer. Subgroup analysis is the best technique to identify and address the issue of varying prediction accuracy across different departments. By analyzing the model's performance for each department separately, the organization can pinpoint if the model is underperforming for specific subgroups. This allows them to identify potential reasons for the discrepancies, such as data imbalances or biases, and take corrective action to improve the model's fairness and accuracy for all departments. **INCORRECT:** "Automated hyperparameter tuning to optimize model accuracy" is incorrect. Hyperparameter tuning can help improve overall model accuracy, but it doesn't specifically address the performance differences across departments. The issue here is more related to bias or data imbalances rather than the overall model accuracy. **INCORRECT:** "Using Amazon Transcribe to improve the training data quality" is incorrect. Amazon Transcribe is a service for converting speech to text and is not relevant to improving the quality of training data for an employee retention model, which likely uses structured data. **INCORRECT:** "Real-time model monitoring with Amazon Polly" is incorrect. Amazon Polly is a service that converts text to speech and is not related to real-time model monitoring or addressing accuracy differences in model predictions across subgroups. **References:** https://docs.aws.amazon.com/sagemaker/latest/dg/model-monitor-data-quality.html Domain: Guidelines for Responsible AI --- #### 62. An agriculture analytics startup collects continuous weather sensor readings (time-series data) and aims to forecast crop yields. Which type of ML technique would be most appropriate for this use case? - Image classification models to detect crop diseases - Regression models to predict crop yields based on sensor data - Reinforcement learning to dynamically adjust irrigation strategies - Clustering on unlabeled data to find sensor anomalies **CORRECT:** "Regression models to predict crop yields based on sensor data" is the correct answer. Regression models are a type of statistical and machine learning model used to analyze relationships between variables and predict continuous numerical values. They help in understanding how dependent variables change concerning independent variables. These models are widely used in finance, healthcare, economics, and engineering for forecasting and data-driven decision-making. Regression is the most appropriate machine learning technique for forecasting crop yields based on continuous weather sensor readings. Since time-series data consists of historical measurements (such as temperature, humidity, and soil moisture), regression models can learn patterns and relationships between these variables and the final crop yield. Common regression techniques for this task include: Linear Regression – If the relationship between sensor data and crop yield is straightforward. Random Forest Regression – Handles nonlinear relationships well and provides robust predictions. Neural Networks (LSTMs for time-series) – Captures temporal dependencies in weather and environmental data. These models help farmers optimize their resources by predicting yields based on current and past environmental conditions. **INCORRECT:** "Clustering on unlabeled data to find sensor anomalies" is incorrect. Clustering is useful for detecting sensor anomalies or segmenting data into groups, but it is not suited for forecasting crop yields, which require a supervised learning approach with labeled data (historical yield values). **INCORRECT:** "Image classification models to detect crop diseases" is incorrect. Image classification models are useful for identifying diseases in crops using satellite or drone images. However, they do not apply to predicting crop yields from time-series weather sensor data. **INCORRECT:** "Reinforcement learning to dynamically adjust irrigation strategies" is incorrect. Reinforcement learning is suitable for optimizing decision-making over time, such as adjusting irrigation strategies based on real-time data. However, it is not the best choice for direct yield prediction, where regression models provide more accurate forecasts. **References:** https://docs.aws.amazon.com/machine-learning/latest/dg/regression.html Domain: Fundamentals of AI and ML --- #### 63. Select and order the following Amazon Bedrock features from the LOWEST to the HIGHEST level of customization required. Each feature should be selected one time. (Select and order THREE.) Note: Select only the correct options, as the type of "Ordering" question is not supported here. - Fine-Tuning - Prompt Engineering - Custom Model Training **CORRECT:** The correct order from lowest to highest customization level is: 1. Prompt Engineering 2. Fine-Tuning 3. Custom Model Training When working with Amazon Bedrock, different levels of customization allow developers to refine AI model behavior according to their needs. Here's the correct order from the lowest to the highest level of customization: Prompt Engineering – This is the easiest and most accessible way to customize a model. It involves crafting well-structured prompts to guide the foundation model's responses without modifying the model itself. Prompt engineering is useful for improving accuracy and relevance in AI-generated outputs without additional training. Fine-Tuning – This method goes a step further by adjusting an existing foundation model with domain-specific data. Fine-tuning allows businesses to improve model performance for specialized tasks, such as legal, medical, or customer service applications, by training it on tailored datasets. Custom Model Training – This is the most advanced and resource-intensive customization level. It involves training a model from scratch using proprietary data, requiring significant computational power and expertise. This option is ideal for organizations that need complete control over their AI model's behavior and performance. Domain: Applications of Foundation Models --- #### 64. A retail company is preparing a dataset with hundreds of features for training a customer behavior model. The team wants to reduce overfitting and improve model interpretability without losing significant predictive power. Which strategy is most suitable? - Use feature transformation to project features into a higher-dimensional space using kernel functions. - Use feature creation to multiply the number of features, enabling the model to capture more interactions. - Apply feature extraction to compress features into embeddings without any label information. - Leverage feature selection to retain only the most relevant predictors for the target variable. **CORRECT:** "Leverage feature selection to retain only the most relevant predictors for the target variable" is the correct answer. Feature selection is a technique used to reduce the number of input features by keeping only those that are most important for predicting the target variable. This helps in reducing overfitting, improving model performance, and making the model easier to interpret. When a dataset contains hundreds of features, many of them might be irrelevant or redundant. By selecting only the most useful features, the model becomes simpler and often performs better. This is especially important in customer behavior models, where interpretability (understanding what drives customer actions) is key. **INCORRECT:** "Use feature creation to multiply the number of features, enabling the model to capture more interactions" is incorrect. Feature creation can be helpful in some cases, but increasing the number of features when you already have hundreds may lead to more overfitting, not less. It also increases complexity, making the model harder to interpret. **INCORRECT:** "Apply feature extraction to compress features into embeddings without any label information" is incorrect. Feature extraction, such as using PCA (Principal Component Analysis), reduces dimensionality but typically loses some interpretability. Also, when done without considering labels (unsupervised), it may not retain the features most relevant to predicting the target variable. **INCORRECT:** "Use feature transformation to project features into a higher-dimensional space using kernel functions" is incorrect. This approach is used in methods like kernel SVMs to find complex patterns. However, projecting into a higher-dimensional space increases complexity and can actually increase the risk of overfitting, which is the opposite of the goal. **References:** https://docs.aws.amazon.com/wellarchitected/latest/machine-learning-lens/feature-engineering.html Domain: Fundamentals of AI and ML --- #### 65. An insurance company is deploying a machine learning model to automatically evaluate claims. Since claim decisions directly affect customers, the legal team is advocating for a model that offers full transparency in its decisions. What are the advantages of using a transparent and explainable model in this scenario? (Select TWO.) - Builds customer trust by showing how claim decisions are made - Guarantees the lowest possible error rate in predictions - Supports regulatory compliance by offering interpretable justifications - Prevents the need for human involvement in the workflow - Increases claim processing speed regardless of data complexity **CORRECT:** "Builds customer trust by showing how claim decisions are made" is a correct answer. Using a transparent and explainable machine learning model helps build customer trust by providing clear insights into how decisions are made. In the insurance industry, customers want to know why their claims are approved or denied. When the model's decision-making process is interpretable, customers feel reassured that the process is fair and unbiased. This transparency strengthens the customer relationship and increases confidence in automated systems. **CORRECT:** "Supports regulatory compliance by offering interpretable justifications" is also a correct answer. In highly regulated industries like insurance, companies must comply with laws that often require them to explain automated decisions. Transparent and explainable models provide clear reasoning for each prediction, making it easier to meet legal requirements. These models can generate human-readable justifications, which can be shared with regulators or used during audits. This not only ensures compliance but also reduces the risk of penalties or reputational damage. **INCORRECT:** "Increases claim processing speed regardless of data complexity" is incorrect. While automation can speed up claims processing, transparency doesn't necessarily affect processing speed. In fact, some explainable models may be slower than black-box models when handling complex data. **INCORRECT:** "Prevents the need for human involvement in the workflow" is incorrect. Explainable models do not eliminate human oversight. In critical applications like insurance claims, human review may still be required to ensure fairness and accuracy, especially for high-stakes decisions. **INCORRECT:** "Guarantees the lowest possible error rate in predictions" is incorrect. Explainable models aim to make decisions understandable—not necessarily to achieve the lowest error rates. In some cases, complex models may be more accurate but less interpretable. **References:** https://docs.aws.amazon.com/sagemaker/latest/dg/clarify-model-explainability.html Domain: Guidelines for Responsible AI