Amazon Practice Questions, Discussions & Exam Topics by our Authors
A company is evaluating several large language models (LLMs) for a text summarization task. The company needs to select a metric to evaluate the quality of the summa...
Let’s carefully analyze each option in the context of evaluating LLM-generated text summaries for quality:
---
A) Recall
What it is: Recall measures the fraction of relevant items that are correctly retrieved (commonly used in classification or information retrieval).
Key factor: Summarization evaluation is not strictly a retrieval problem; it’s about text similarity, content coverage, and linguistic quality, not just capturing “all relevant items.”
Scenario where useful: Recall is appropriate in information retrieval, search engines, or classification tasks where you want to ensure most relevant items are included.
Why rejected here: It does not capture the nuances of summary quality or textual similarity to reference summaries.
---
B) Area under the ROC curve (AUC)
What it is: AUC measures the ability of a classifier to distinguish between classes.
Key factor: Summarization is not a binary classification task, so AUC is irrelevant.
Scenario where useful: Binary or multi-class classification evaluation, e.g., predicting whether an email is spam.
Why rejected here: It provides no meaningful evaluation for text content or fluency in summaries.
---
C) Recall-Oriented Understudy for Gisting Evaluation (ROUGE)
What it is: ROUGE is a text similarity metric d...
Author: Isabella · Last updated Jul 19, 2026
A research group wants to test different generative AI models to create research papers. The research group has defined a prompt and needs a method to assess the models' output. The research group wants to use a team of...
Let’s carefully go through the options and reason which one fits the scenario. The key points from your scenario are:
The goal is to test different generative AI models creating research papers.
There is a defined prompt.
Human scientists will assess the outputs.
This is specifically in the AWS ecosystem.
Now, analyzing each option:
---
A) Use automatic evaluation on Amazon Personalize
Amazon Personalize is designed for personalized recommendations, like suggesting products or content to users.
It is not intended for evaluating AI-generated text or research papers.
Rejection reasoning: The key factor here is that Personalize cannot perform output assessments of generative AI.
---
B) Use content moderation on Amazon Rekognition
Amazon Rekognition is for image and video analysis, including detecting unsafe content, faces, objects, and moderation.
It cannot evaluate textual research papers.
Rejection reasoning: Rekognition is irrelevant because the task is text-based, not images/videos.
---
C) Use model evaluation o...
Author: ElectricLionX · Last updated Jul 19, 2026
SNAPSHOT
-
An ecommerce company is developing a generative AI solution to create personalized product recommendations for its application users. The company wants to track how effectively the AI solution increases product sales and user engagement in the application.
Select th...
Author: NightmareDragon2025 · Last updated Jul 19, 2026
An AI practitioner wants to evaluate ML models. The AI practitioner wants to provide explanations of model predictions to customers and stakeholder...
Let’s carefully go through each option and reason which one fits the requirements. The scenario is:
Requirements:
Evaluate ML models
Provide explanations of model predictions to customers/stakeholders
---
Option A: Amazon QuickSight
What it does: Business intelligence (BI) service for creating dashboards, visualizations, and analytics.
Evaluation: QuickSight is excellent for visualizing data and trends, but it does not provide explanations of ML model predictions. It’s primarily for reporting and dashboards.
Conclusion: ❌ Not suitable for explaining ML predictions.
---
Option B: Amazon Comprehend
What it does: NLP service for text analytics—e.g., sentiment analysis, entity recognition, and topic modeling.
Evaluation: Useful for analyzing text data, but it does not evaluate arbitrary ML models or explain predictions.
Conclusion: ❌ Not suitable.
---
Option C: AWS Trusted Advisor
What it does: Provides recommenda...
Author: ShadowWolf101 · Last updated Jul 19, 2026
Sentiment analysis is a subset of which broader field of AI?
Let’s carefully break this down. The question is: “Sentiment analysis is a subset of which broader field of AI?”
We’ll analyze each option:
---
A) Computer Vision
What it is: Deals with understanding and interpreting visual data like images and videos.
Relevance to sentiment analysis: Sentiment analysis focuses on text (or sometimes speech) to determine emotions or opinions. It doesn’t primarily process images or video.
Scenario where used: Detecting objects in photos, facial recognition, autonomous driving.
Conclusion: Not a match.
---
B) Robotics
What it is: AI applied to physical machines or robots to interact with the real world.
Relevance to sentiment analysis: Robotics may use AI for perception or decision-making, but sentiment analysis is not about controlling robots.
Scenario where used: Robot navigation, warehouse automation, robotic arms.
Conclusion: Not a match.
---
C) Natural Language Processing (NLP)
What it is: AI field that deals with understanding, interpreting, and generating human language.
Re...
Author: Ming · Last updated Jul 19, 2026
A company wants to set up private access to Amazon Bedrock APIs from the company's AWS account. The company also wants to protect its data fro...
Let’s carefully analyze this AWS scenario step by step. The requirements are:
1. Private access to Amazon Bedrock APIs from the company’s AWS account.
2. Protect company data from internet exposure.
We’ll evaluate each option based on these requirements.
---
Option A: Use Amazon CloudFront to restrict access to the company’s private content
CloudFront is a content delivery network (CDN) primarily used for serving web content with low latency.
While CloudFront can restrict access to content via signed URLs or geo-restrictions, it does not provide private API access or a secure direct connection to AWS services like Bedrock.
Not suitable for this scenario because it’s meant for content distribution, not private API connectivity.
Rejected.
---
Option B: Use AWS Glue to set up data encryption across the company’s data catalog
AWS Glue is an ETL (Extract, Transform, Load) service for managing and preparing data.
It can encrypt data in catalogs or during processing, but it does not create private connections to APIs.
Encrypting data alone does not prevent exposure of the API traffic to the internet.
Rejected.
---
Option C: Use AWS Lake Formation to manage cen...
Author: StarlightBear · Last updated Jul 19, 2026
A company receives a large amount of unstructured user feedback in text format. The company wants to analyze the sentiment of the us...
Let’s carefully evaluate each option based on the problem requirements: analyzing sentiment from unstructured user feedback in text.
---
Option A: Use a large language model (LLM) to perform natural language processing (NLP) for sentiment analysis ✅
Reasoning:
LLMs (like Amazon Bedrock models or SageMaker integrated models) are designed for understanding unstructured text.
They can perform sentiment analysis directly, identifying positive, negative, or neutral sentiment.
They work well with large volumes of text data, which matches the requirement.
Scenario fit: When you have free-text user feedback and need accurate sentiment classification.
---
Option B: Use a regression algorithm to classify the feedback based on predefined categories, then analyze sentiment ❌
Reasoning:
Regression is mainly for predicting numeric outcomes, not categorical sentiment labels.
Classifying text first would require feature engineering (like converting text to vectors) and then mapping categories to sentiment, which is more complex and indirect than using NLP directly.
Scenario fit: More suitable if you are predicting quantitative outcomes from structured features, not raw text sentime...
Author: Isabella · Last updated Jul 19, 2026
SNAPSHOT
-
A company wants to improve multiple ML models.
Select the correct technique from the following list of use cases. ...
Author: Scarlett · Last updated Jul 19, 2026
A company wants to create an AI solution to generate images and descriptions for a product catalog. The company needs to select a foundation model (FM) for this solution.
The company must consid...
The company is evaluating which types of outputs a foundation model can produce (images and text descriptions). In AWS terminology, this characteristic is modality.
Let’s break down each option, why it fits or doesn’t fit, and when it would be the right choice.
---
✅ Correct Answer: D) Modality
Why this is correct
Modality refers to the type of input and output data a foundation model supports, such as:
Text → Text
Text → Image
Image → Text
Multimodal (text, image, audio, video)
The company needs:
Generated images (product visuals)
Generated descriptions (text)
Therefore, they must choose an FM that supports both image and text generation, which is a multimodal model (for example, models like Amazon Titan Image Generator or multimodal models in Amazon Bedrock).
Key factor
Output type is the primary concern, not speed, size, or tuning.
Modality directly answers: “Can this model generate the kinds of outputs we need?”
Typical scenario
Product catalogs
Marketing content
E-commerce listings
Any solution requiring text + images
---
❌ A) Latency
Why it’s rejected
...
Author: Abigail · Last updated Jul 19, 2026
A company wants to use an ML model to analyze customer reviews on social media. The model must determine if each review has a neutral, positive, or negative senti...
Correct approach: D) Classification
Why Classification fits the requirement
The company needs the ML model to assign each customer review to exactly one of three predefined labels: positive, neutral, or negative.
This is a textbook text classification problem.
Key factors in the reasoning:
Discrete output labels: The sentiment categories are fixed and known in advance.
One-to-one mapping: Each review maps to a single sentiment class.
Measurable evaluation: Classification enables clear metrics such as accuracy, precision, recall, F1 score, which are commonly used in AWS services like Amazon Comprehend.
Scalability: Classification models are efficient and well-suited for large volumes of social media text.
AWS context:
AWS typically solves sentiment analysis using classification models, for example:
Amazon Comprehend’s Sentiment Analysis API
Custom text classification models trained wit...
Author: Stella · Last updated Jul 19, 2026
SNAPSHOT
-
Select the correct AI term from the following list for each statement. Each AI term s...
Author: Rahul · Last updated Jul 19, 2026
Which option is an example of unsupervised learning?
The correct example of unsupervised learning is Option A. Here’s the breakdown using key ML factors (especially how AWS frames these concepts for exams).
---
Key factor to identify unsupervised learning
Unsupervised learning works with unlabeled data. The model is not told the “right answer” ahead of time; instead, it discovers patterns, structure, or relationships on its own.
---
Option-by-option explanation
A) Clustering data points into groups based on their similarity ✅
Why this is correct:
No labels are provided (no predefined group names).
The algorithm finds patterns and groups data based on similarity.
This is a classic unsupervised task.
Typical scenarios where this is used:
Customer segmentation (grouping customers by behavior)
Anomaly detection
Market basket analysis
AWS examples: Amazon SageMaker K-means, anomaly detection services.
---
B) Training a model to recognize images of animals ❌
Why this is rejected:
Requires labeled image...
Author: Ella · Last updated Jul 19, 2026
An online learning company with large volumes of education materials wants to use enterprise search.
...
Correct answer: Amazon Kendra
Why Amazon Kendra is selected
An online learning company with large volumes of education materials needs enterprise search—a service that can index, understand, and retrieve information across many documents (PDFs, Word files, FAQs, presentations, websites, etc.).
Amazon Kendra is purpose-built for this use case.
Key factors:
Enterprise-grade search: Designed specifically for searching large document repositories.
Natural language understanding: Users can ask questions in plain English (e.g., “What is photosynthesis?”) and get precise answers.
Semantic relevance: Goes beyond keyword matching to understand intent and context.
Built-in connectors: Easily connects to data sources like S3, SharePoint, databases, and learning management systems.
Scalability: Handles large volumes of educational content efficiently.
This makes Amazon Kendra ideal for students, instructors, or employees searching through massive learning content libraries.
---
Why the other options are rejected
A) Amazon Comprehend ❌
W...
Author: Sofia2021 · Last updated Jul 19, 2026
A company creates video content. The company wants to use generative AI to generate new creative content and to reduce video creation time.
Which solution ...
Best choice: Option C — Use the Amazon Nova Reel model on Amazon Bedrock to generate videos.
Below is a structured explanation of why Option C is the most operationally efficient, and why the other options are less suitable, using key decision factors.
---
Key requirements from the scenario
1. Generate new creative content → Needs a generative AI model designed for creative media.
2. Reduce video creation time → Minimize manual steps and post-processing.
3. Most operationally efficient → Fewer services, fewer workflows, lower integration and maintenance overhead.
---
Option-by-option analysis
Option A
Use Amazon Titan Image Generator on Bedrock + video editing software
Why it’s rejected
Titan Image Generator is image-only, not video-native.
Requires a multi-step workflow:
Generate images
Export images
Manually assemble and animate them in video software
High operational overhead and longer production timelines.
When this option makes sense
Static image generation (storyboards, thumbnails, marketing graphics)
Teams that already rely heavily on manual video editing and only want AI-assisted visuals
Key limitation
❌ Does not significantly reduce video creation time
---
Option B
Use Amazon Nova Canvas on Bedrock + video editing software
Why it’s rejected
Nova Canvas is optimized for high-quality image generation and design tasks, not video creation.
Still requires external video tools and manual sequencing.
Similar operational inefficiencies as Option A, despite better creative image quality.
When this option makes sense
Visual design workflows (concept art, backgrounds, keyframes)
Pre-production assets rather than final video output
Key limitation
❌ Image-first model, not video-first
---
Option C (Selected)
...
Author: NightmareDragon2025 · Last updated Jul 19, 2026
A company is training ML models on datasets. The datasets contain some classes that have more examples than other classes. The company wants to measure how well the model balance...
The key issue in this scenario is class imbalance (some classes have many more examples than others) and the need to balance detecting classes and labeling them correctly. This wording is very important for choosing the metric.
---
Correct Answer: D) F1 score
Why F1 score is the best choice
The F1 score is the harmonic mean of Precision and Recall. It is designed specifically for situations where:
Classes are imbalanced
You care about both:
Detecting as many relevant examples as possible (Recall)
Making sure predictions are correct when the model predicts a class (Precision)
Key factors in the reasoning:
Accuracy can be misleading with imbalanced data
Precision alone ignores missed detections
Recall alone ignores false positives
F1 score balances both types of errors
In AWS ML and exam scenarios, F1 score is the standard choice when the question mentions class imbalance and balance between detection and correctness.
---
Why the other options are rejected
A) Accuracy ❌
Why rejected:
Accuracy assumes balanced classes
With imbalanced data, a model can get very high accuracy by always predicting the majority ...
Author: Sam · Last updated Jul 19, 2026
A company is analyzing financial transaction records. The company categorizes the records as either personal or business. The company inserts the categories into the...
This scenario describes adding category information (personal or business) directly into transaction records so that the data can be identified and used correctly later (for example, for analytics or machine learning).
Let’s evaluate each option using key data-preparation factors and explain why one fits and the others do not, in an AWS context.
---
Correct Option
B) Data labeling ✅
Why this is correct:
Data labeling is the process of assigning meaningful tags or categories to data.
In this scenario, the company categorizes transactions as personal or business and inserts those categories into the records.
This is exactly what labeling means: adding human-defined or business-defined labels to data.
AWS context example:
In Amazon SageMaker, data labeling is used to prepare datasets for supervised learning.
A transaction dataset labeled as personal or business can later be used for:
Fraud detection
Spending pattern classification
Cost allocation models
Key factor:
The data itself is unchanged in value or scale; it is simply annotated with a category.
---
Why the Other Options Are Incorrect
A) Data encoding ❌
What it is:
Converts data from one format to another (e.g., text ...
Author: Nia · Last updated Jul 19, 2026
A company wants to extract key insights from large policy documents to increase employee efficiency.
Which...
Correct strategy: C) Summarization
Why Summarization is selected
The requirement is to extract key insights from large policy documents to increase employee efficiency. This aligns directly with text summarization, a generative AI capability commonly used in AWS services (for example, Amazon Bedrock with foundation models or Amazon Comprehend for extractive summaries).
Key factors in the reasoning:
Unstructured text input: Policy documents are long, text-heavy, and unstructured.
Goal is condensation, not prediction: Employees need shorter, meaningful versions of documents, not numerical outputs.
Generative behavior: Summarization generates new text that captures the most important points.
Efficiency improvement: Summaries reduce reading time while preserving essential information.
Typical AWS scenario:
Automatically generating executive summaries of HR, legal, or compliance policies so employees can quickly understand updates.
---
Why the other options are rejected
A) Regression — ...
Author: Liam · Last updated Jul 19, 2026
A company is using Amazon SageMaker to deploy a model that identifies if social media posts contain certain topics. The company needs to show how different input features inf...
Let’s carefully go through the options and determine which SageMaker feature meets the requirement: showing how different input features influence model behavior.
---
A) SageMaker Canvas
Purpose: Enables business analysts to build ML models without writing code. Focuses on model creation and predictions in a no-code interface.
Key factor: Canvas is mainly for building models easily, not for explaining model behavior or feature influence.
Scenario: Useful when non-technical users need to quickly create models and get predictions.
Verdict: ❌ Not suitable for explaining feature influence.
---
B) SageMaker Clarify
Purpose: Detects bias in models and datasets and provides explainability for ML models.
Key factor: Clarify can generate feature importance explanations, showing how each input feature contributes to predictions. It can work for tabular, image, or text data.
Scenario: Perfect when a company wants to show how input features impact model behavior.
...
Author: Olivia · Last updated Jul 19, 2026
SNAPSHOT
-
An AI practitioner is determining the appropriate data type for various use cases.
Select the correct data type from the fol...
Author: Nathan · Last updated Jul 19, 2026
A company wants to assess internet quality in remote areas of the world. The company needs to collect internet speed data and store the data in Amazon RDS. The company will analyze internet speed variation throughout each day. The company wants to create an AI ...
Let’s break this down carefully step by step. The company wants to assess internet quality in remote areas, analyze internet speed variations throughout the day, and predict potential internet disruptions.
We need to consider the type of data that fits these requirements and how AWS tools handle them.
---
Step 1: Analyze the scenario
Data to collect: Internet speed (e.g., download/upload speed, latency) over time.
Storage: Amazon RDS (relational database service), which works well for structured data.
Analysis goal: See how internet speed changes during the day.
Prediction goal: AI model to predict disruptions based on trends.
Key factors:
1. Time-dependency: Internet speed varies over time → time is crucial.
2. Structured measurements: Download speed, upload speed, latency, timestamp → numeric values associated with a time index.
3. Predictive modeling: We need patterns in temporal data to forecast future internet disruptions.
---
Step 2: Evaluate the options
A) Tabular data
What it is: Structured data stored in rows and columns (e.g., CSV, SQL tables).
Pros: Can store speed, latency, location, timestamp.
Cons: Tabular data alone doesn’t capture the temporal sequence inherently; you have to manage time explicitly.
Use case: Works for generic structured data (e.g., customer info, product inventory) but not ideal if we want to predi...
Author: Zain · Last updated Jul 19, 2026
A company wants to build an ML model to detect abnormal patterns in sensor data. The company does not have labeled data for tra...
Let’s carefully analyze this scenario step by step.
Scenario:
The company wants to detect abnormal patterns in sensor data.
No labeled data is available for training.
The goal is anomaly detection in AWS (or in general ML terms).
---
Option Analysis:
A) Linear Regression
Linear regression predicts a continuous target variable from input features.
Key factor: Requires labeled data (target variable).
Relevance to scenario: The company has no labels, and detecting anomalies is not about predicting a numeric value—it’s about identifying unusual patterns.
Conclusion: ❌ Not suitable.
---
B) Classification
Classification predicts discrete labels (e.g., “normal” vs “abnormal”).
Key factor: Requires labeled training data with predefined classes.
Relevance to scenario: The company has no labels, so it cannot train a classifier.
Conclusion: ❌ Not suitable.
---
C) Decision Tree
Decision trees are supervised learning algorithms. They can be used for regression or classification.
Key factor: Requires labeled data.
Relevance to scenario: Same issue as classificat...
Author: Ethan Smith · Last updated Jul 19, 2026
A company uses Amazon Bedrock to implement a generative AI assistant on a website. The AI assistant helps customers with product recommendations and purchasing decisions.
The company wants to measure the direct i...
Let’s carefully analyze each option based on the goal: measuring the direct impact of the AI assistant on sales performance.
---
A) The conversion rate of customers who purchase products after AI assistant interactions ✅
Why it fits: Conversion rate directly measures how many users complete a purchase after interacting with the AI assistant. This directly links the AI’s performance to sales outcomes, which is exactly what the company wants.
When to use: Anytime you want to measure the financial impact or effectiveness of AI in driving purchases.
---
B) The number of customer interactions with the AI assistant ❌
Why it’s not sufficient: The number of interactions only tells you how many times the AI was used, not whether it led to actual sales. High interactions don’t necessarily translate to higher revenue.
When to use: Useful for tracking engagement or usage, but not direct sales impact.
---
C) Sentiment analysis scores from customer feedback after AI assistant interactions...
Author: Noah · Last updated Jul 19, 2026
Which AWS service or feature stores embeddings in a vector database for use with foundation models (FM...
Let’s carefully analyze your question. You’re asking about storing embeddings in a vector database for use with foundation models (FMs) and Retrieval-Augmented Generation (RAG) in AWS. We’ll go option by option.
---
A) Amazon SageMaker Ground Truth
Purpose: This is a data labeling service for machine learning. It helps create labeled datasets for training models (e.g., image labeling, text classification).
Relevance to embeddings/vector storage: ❌ Not relevant. Ground Truth does not store embeddings or act as a vector database. It’s strictly for creating labeled datasets before training.
Scenario: Use when you need human-labeled training data for supervised ML models.
---
B) Amazon OpenSearch Service
Purpose: Managed search and analytics service. It now supports k-NN vector search, which allows storing vector embeddings for similarity search.
Relevance to embeddings/vector storage: ✅ Highly relevant. You can store embeddings and perform fast similarity searches, which is exactly what RAG workflows require (retrieving relevant context from a knowledge base for LLMs).
Scenario: Use OpenSearch when you want to store document embeddings and retrieve similar vectors to feed into an LLM for RAG. Key factors: ve...
Author: Liam · Last updated Jul 19, 2026
Which scenario represents a practical use case for generative AI?
Let’s carefully analyze each option using key factors about generative AI and its practical use cases:
---
A) Using an ML model to forecast product demand
Analysis: This is a predictive task. Traditional machine learning models (like regression or time-series forecasting) are ideal here.
Generative AI fit: Low. Generative AI is designed to create new content (text, images, code, etc.), not primarily to forecast numeric demand.
Conclusion: Not the best fit.
---
B) Employing a chatbot to provide human-like responses to customer queries in real time
Analysis: Generative AI excels at producing human-like text in real time. Chatbots powered by models like GPT can generate responses that feel natural, personalized, and context-aware.
Generative AI fit: High. This is a classic and practical use case.
Conclusion: Strong candidate.
---
C) Using an analytics dashboard to track website traffic and user behavior
Analysis: This is a descriptive/monitoring task. Dashboards aggregate and visualize data but don’t generate new content.
...
Author: Kai · Last updated Jul 19, 2026
A company is using Amazon Bedrock for a generative AI solution. The solution must integrate a service with vector database storage and vector search ...
Let’s break this down carefully. The requirement is:
Generative AI solution using Amazon Bedrock
Needs integration with vector database storage
Requires vector search capabilities
We’ll analyze each option against these key factors.
---
A) Amazon DynamoDB
What it is: NoSQL key-value and document database.
Strengths: Low-latency storage for structured data, highly scalable.
Limitations: DynamoDB is not designed for vector storage or vector similarity search out of the box. You’d need extra custom logic to handle embeddings and similarity queries, which is not native.
Scenario fit: Great for session data, caching metadata, or structured key-value workloads.
Conclusion: Rejected because vector search is a core requirement, which DynamoDB does not natively support.
---
B) Amazon OpenSearch Service
What it is: Managed search and analytics engine (successor to Elasticsearch).
Strengths: Supports vector search (kNN search), full-text search, and analytics. Can store embeddings generated by generative AI models and perform similarity searches efficiently.
Scenario fit: Ideal for applications where generative AI outputs (e.g., embeddings) need fast retrieval based on similarity. Can integrate directly with Bedrock for semantic search workflows.
Conclusion: Strong fit because it meets both vector storage and vector search requirements.
---
C) Amazon ElastiCache
What it is:...
Author: Max · Last updated Jul 19, 2026
A media streaming platform wants to provide movie recommendations to users based on the users' account history...
Let's carefully analyze each option based on the scenario: recommending movies to users based on their account history.
---
Scenario requirement:
Provide personalized movie recommendations based on user behavior/history.
Key factors: personalization, machine learning recommendations, user-specific data.
---
Option A: Amazon Polly
Purpose: Converts text into natural-sounding speech.
Use case: Voice assistants, reading content aloud.
Why rejected: Polly is for text-to-speech, not recommendations or analyzing user behavior.
---
Option B: Amazon Comprehend
Purpose: Natural Language Processing (NLP) service that detects sentiment, key phrases, entities, and language in text.
Use case: Analyzing user reviews, extracting insights from text.
Why rejected: While it can analyze...
Author: Aarav · Last updated Jul 19, 2026
A company has developed an ML model to approve or reject loan applications. The model's decision-making process must be transparent and explainable to comply with regulatory requirements. The company must document th...
Let’s carefully analyze each option in the context of the problem.
Scenario:
A company has an ML model for loan approvals.
Requirement: The model must be transparent and explainable for regulatory compliance.
Requirement: The company must document the decision-making process for audits.
We need a solution that helps with ML model explainability and documentation, not general data extraction or infrastructure provisioning.
---
Option A: Amazon Textract
Purpose: Extracts text and data from scanned documents or images.
Relevance: Textract is used for document processing, not ML model explainability.
Verdict: ❌ Rejected. It does not provide insight into how an ML model makes decisions.
---
Option B: Amazon SageMaker Model Card
Purpose: Provides a standardized way to document ML models, including:
Intended use
Performance metrics
Limitations and biases
Ethical considerations and compliance notes
Relevance: Directly addresses regulatory documentation and transparency, helping auditors understand the ML model’s decision-making.
Scenario: Use SageMaker Model Cards whenever you need to document a m...
Author: Emma Brown · Last updated Jul 19, 2026
SNAPSHOT
-
A company is building a generative AI application and is reviewing foundation models (FMs). The company needs to consider multiple FM characteristics.
Select the correct FM characteristic from ...
Author: Ishaan · Last updated Jul 19, 2026
A company is using large language models (LLMs) to develop online tutoring applications. The company needs to apply configurable safeguards to the LLMs. These safeguards must ensure that the LLMs follow standard safety rules ...
Let's carefully analyze this AWS scenario:
Scenario:
A company is developing online tutoring applications using LLMs.
They need configurable safeguards to ensure LLMs follow standard safety rules.
Goal: least effort to implement.
We are asked to evaluate the AWS options:
---
Option A: Amazon Bedrock Playgrounds
Purpose: Bedrock Playgrounds is a web-based interface to experiment with Foundation Models (LLMs) without writing code.
Pros: Easy for testing models.
Cons: It is only for experimentation, not for applying configurable safety safeguards in production.
Verdict: ❌ Not suitable for enforcing safety rules.
---
Option B: Amazon SageMaker Clarify
Purpose: SageMaker Clarify is used for bias detection and explainability in ML models.
Pros: Helps detect bias and explain predictions for models like tabular, text, or images.
Cons: It does not apply runtime safeguards to LLMs. It’s more about auditing than controlling behavior.
Verdict: ❌ Not suitable for applying safety rules in LLM applications.
---
Option C: Amazon Bedrock Guardrails
Purpose: Bedrock Guardrails allo...
Author: Emily · Last updated Jul 19, 2026
A company is exploring Amazon Nova models in Amazon Bedrock. The company needs a multimodal model that supports multiple languages.
Which N...
Here’s a clear AWS‑focused explanation of which Amazon Nova model in Amazon Bedrock meets the requirement of multimodal + multilingual most cost‑effectively, with reasoning on all options:
---
✅ Key Requirement
Multimodal → supports text, images, and video inputs (not just text). ([Amazon Web Services, Inc.][1])
Multiple languages → supports 200+ languages is expected for understanding models. ([Channel Drive][2])
Cost‑effective → lower cost while still meeting the core task needs.
---
Option Analysis
A) Nova Lite
✔️ Multimodal (text, image, video)
✔️ Multilingual (supports 200+ languages) ([Amazon Web Services, Inc.][1])
💡 Cost‑effective — designed as a very low‑cost multimodal option. ([Amazon Web Services, Inc.][1])
Best fit if: the company needs a budget‑friendly multimodal model that supports languages and basic general tasks.
Why choose over others: it’s cheaper than Pro but still handles multimodal + multilingual.
When it may fall short: not as capable/accurate as Pro for advanced or highly complex tasks.
➡️ Strong choice for balancing cost with multimodal & multilingual support.
---
B) Nova Pro
✔️ Multimodal (text, image, video)
✔️ Multilingual (supports 200+ languages) ([Amazon Web Services, Inc.][1])
🟡 Higher capability (better accuracy/complex reasoning) but higher cost than Lite. ([Amazon Web Services, Inc.][1])
Best fit if: the company needs more accuracy, reasoning, or complex multimodal tasks.
Why not the top answer: it is not the most cost‑effecti...
Author: Emma · Last updated Jul 19, 2026
A company is building a new generative AI chatbot. The chatbot uses an Amazon Bedrock foundation model (FM) to generate responses. During testing, the company notices that the chatbot is prone to prompt injection at...
Let’s carefully analyze each option based on effectiveness, implementation effort, and the scenario:
---
Scenario: A generative AI chatbot on Amazon Bedrock is vulnerable to prompt injection attacks. The company wants to secure it with minimal effort.
---
Option A: Fine-tune the FM to avoid harmful responses
Pros: Fine-tuning can customize the model’s behavior to avoid certain types of harmful outputs.
Cons: Fine-tuning requires significant effort, including data preparation, training, testing, and deployment.
Evaluation: Effective but high implementation effort, not the least effort solution.
Scenario suitability: Use when you need deep, persistent customization beyond basic safety filters.
---
Option B: Use Amazon Bedrock Guardrails content filters and denied topics
Pros:
Guardrails allow you to define rules for denied topics, prohibited content, and safety filters.
Minimal implementation effort: you configure rules instead of retraining the model.
Specifically designed to mitigate prompt injection and unsafe content in Bedrock.
Cons: Less flexible than fine-tuning for highly specialized behaviors, but sufficient for general safety.
Evaluation: Best fit for “least effort” security solution.
Scenario suitabilit...
Author: Alexander · Last updated Jul 19, 2026
What does inference refer to in the context of AI?
Let’s carefully analyze the options in the context of AI inference, especially as used in AWS:
---
Option A: The process of creating new AI algorithms
Analysis: This describes model development or research, where new algorithms are designed.
Why rejected: Inference does not involve creating algorithms; it uses already trained models.
Scenario where this applies: Research teams designing novel neural network architectures.
---
Option B: The use of a trained model to make predictions or decisions on unseen data
Analysis: This is exactly what inference means. Once a model is trained (on training data), inference is when the model is deployed to predict outcomes or classify new, unseen inputs.
Key factors:
Requires a pre-trained model.
Works on new or unseen data.
Can be done in real-time (e.g., AWS SageMaker endpoints) or batch mode...
Author: RadiantJaguar56 · Last updated Jul 19, 2026
A company wants to build an AI assistant to provide responses to user queries. The AI assistant must evaluate specific data sources, query external APIs, generate response options, and compare and prioritize r...
Let's carefully analyze the question. The company wants an AI assistant that can:
1. Evaluate specific data sources.
2. Query external APIs.
3. Generate response options.
4. Compare and prioritize those response options.
Now, let's evaluate the options in Amazon Bedrock:
---
A) Prompt Management
What it does: Helps structure and manage prompts sent to a foundation model.
Why it doesn’t fit: Prompt management focuses only on constructing prompts and managing their versions. It doesn’t handle external API calls, evaluating data sources, or comparing multiple response options.
Scenario for use: When you want to improve prompt quality or manage prompt templates.
---
B) Response Streaming
What it does: Streams partial responses from a model in real time.
Why it doesn’t fit: This is purely about delivering responses faster, not about evaluating data, querying APIs, or prioritizing answers.
Scenario for use: When you need low-latency responses or streaming output in real-time applications.
---
C) Knowledge Bases
What it does: Stores structured or unstructured data for a model t...
Author: Alexander · Last updated Jul 19, 2026
An AI practitioner notices a large language model (LLM) is generating different responses for the same input across multipl...
Let’s carefully analyze this AWS AI question step by step.
Scenario: An AI practitioner notices that a large language model (LLM) generates different responses for the same input across multiple invocations.
We are asked to identify which AI risk this describes, explain why the correct option fits, and why the others do not.
---
Option Analysis
A) Hallucinations
Definition: Hallucinations occur when an AI model generates outputs that are factually incorrect, misleading, or nonsensical.
Key factors: Focus is on accuracy of content, not variability of outputs.
Why it’s rejected: The question is about different responses for the same input, not whether those responses are factually correct or not. Hallucination is about truthfulness, not consistency.
B) Nondeterminism
Definition: Nondeterminism refers to the property of an AI model where it can produce different outputs given the same input, usually due to sampling methods, randomness in the model, or temperature settings.
Key factors: Directly matches the scenario described. LLMs often use stochastic sampling, s...
Author: Rahul · Last updated Jul 19, 2026
A company is building a generative AI application on AWS. The application will help improve reading comprehension for students. The application must give students the ability to...
Let's carefully analyze the options based on the requirement: students should be able to add illustrations to stories in a generative AI application on AWS. The key factors here are:
Illustration generation (image creation from text)
Integration with generative AI
Focus on student reading comprehension (enhanced story experience)
---
Option A: Use Amazon Bedrock Stable Diffusion 3.5 Large to generate images based on text inputs
Analysis: Stable Diffusion is a text-to-image model, perfect for generating illustrations from story text.
Pros: Directly fulfills the requirement of adding illustrations. Bedrock lets you use foundation models without managing infrastructure.
Conclusion: ✅ This is suitable for generating illustrations dynamically from the students’ text input.
---
Option B: Use Amazon Polly to create an audiobook based on story texts
Analysis: Amazon Polly converts text into speech, creating audio versions of the story.
Pros: Good for listening comprehension.
Cons: Does ...
Author: Kai · Last updated Jul 19, 2026
A healthcare company wants to analyze patient data. The data was gathered over the previous year to detect patterns in disease outbreaks. The company needs to create a trend analysis report for each month to present to public health officials. The company must provide insights into patient dat...
Let’s carefully analyze the scenario and each AWS inference option.
Scenario requirements:
1. Analyze historical patient data (from the past year) to detect patterns.
2. Create monthly trend analysis reports.
3. Provide insights from the most recent month.
4. Must be cost-effective.
Key points:
The data is already collected, so real-time streaming isn’t required.
Insights are produced periodically, likely once a month, not continuously.
Cost efficiency is important.
---
Option analysis
A) Real-time inference
Definition: Model predicts immediately for each incoming request.
Use case: Ideal for live, low-latency predictions (e.g., recommending products in real-time or fraud detection).
Reason rejected: Patient data is batch historical data, not streaming. Real-time endpoints incur always-on costs, so not cost-effective for periodic analysis.
---
B) Batch Transform
Definition: Processes large datasets at once, running inference in bulk.
Use case: Best for historical datasets or periodic analysis where low latency is not needed.
Advantages here:
Can process all patient records for a month in one job.
Cost-efficient since you pay per job, no always-on endpoint.
Supports large datasets, which is typical for patient records ...
Author: ShadowWolf101 · Last updated Jul 19, 2026
SNAPSHOT
-
Select and order the steps from the following list to correctly describe the ML lifecycle for...
Author: William · Last updated Jul 19, 2026
A company acquires International Organization for Standardization (ISO) accreditation to manage AI risks and to use AI responsibly.
...
Let's analyze this carefully. The question is about a company acquiring ISO accreditation to manage AI risks and use AI responsibly, specifically in an AWS context. The goal is to determine what this accreditation actually reflects about the company.
---
Option Analysis
A) All members of the company are ISO certified
Reasoning: ISO accreditation is usually granted to the organization or its processes, not to individual employees. While employees may receive training or certification in certain ISO standards, the company as a whole doesn’t imply that every employee is ISO certified.
Conclusion: Incorrect for this scenario.
B) All AI systems that the company uses are ISO certified
Reasoning: ISO accreditation generally focuses on management processes, risk frameworks, and governance, not on certifying individual AI systems. AI systems themselves are not “ISO certified”; the company’s processes for designing, deploying, or managing AI can be certified.
Conclusion: Incorrect.
C) All AI application team members are ISO certified
Reasoning: Similar to option A, ISO accreditation is process-focused, not a measure of individual team certifications. While team members may follow ISO-aligned processes, ISO does not certify people as part of...
Author: StarryEagle42 · Last updated Jul 19, 2026
SNAPSHOT
-
Select the correct prompt engineering technique from the following list for each description. Select each ...
Author: Max · Last updated Jul 19, 2026
A company is developing an ML model to predict heart disease risk. The model uses patient data, such as age, cholesterol, blood pressure, smoking status, and exercise habits. The dataset includes a target value that ind...
Let’s carefully analyze the scenario and the options:
Scenario details:
The company wants to predict heart disease risk.
The dataset includes patient features (age, cholesterol, blood pressure, smoking status, exercise habits).
The dataset also includes a target value: whether the patient has heart disease (yes/no).
From this, we can identify key factors:
1. Target label exists → This is crucial. It means the data is labeled.
2. Goal is prediction → We want the model to learn patterns from features to predict the target.
---
Option analysis:
A) Unsupervised learning
Works with unlabeled data. Examples: clustering, anomaly detection.
Key use: Finding patterns or groupings without known outcomes.
Rejection reason: In this scenario, we already have the target label (heart disease yes/no). So unsupervised learning is not appropriate here.
B) Supervised learning
Works with labeled data, where the model learns the mapping from features → target.
Key use: ...
Author: ShadowWolf101 · Last updated Jul 19, 2026
SNAPSHOT
-
A company periodically updates its product database by manually uploading digital product guides. The product guides contain text and images. The company wants to automate this task by using generative AI.
Select and order the steps fr...
Author: CrimsonViperX · Last updated Jul 19, 2026
A company has guidelines for data storage and deletion.
Which data governance strategy does this ...
Let’s carefully analyze this AWS data governance question step by step.
Question: A company has guidelines for data storage and deletion. Which data governance strategy does this describe?
Step 1: Examine the options
A) Data de-identification
Definition: Removing or masking personally identifiable information (PII) so that data cannot be traced back to an individual.
Key factor: Focuses on privacy and anonymization, not on storage or deletion rules.
Rejection reasoning: The scenario is about storage and deletion, not anonymization.
B) Data quality standards
Definition: Ensures that data is accurate, complete, and consistent.
Key factor: Focuses on correctness and reliability of data.
Rejection reasoning: Guidelines for storage or deletion are unrelated to data accuracy or consistency.
C) Data retention
Definition: Po...
Author: Ahmed97 · Last updated Jul 19, 2026
A company needs to apply numerical transformations to a set of images to transpose and rotate the images.
Which solution will meet t...
Let’s carefully analyze the options and reasoning based on AWS services and the requirement: apply numerical transformations (transpose and rotate) to images in an operationally efficient way.
---
Option A: Create a deep neural network using the images as input
Pros: Neural networks can learn complex transformations.
Cons:
Overkill for simple image operations like transpose and rotation.
Requires training data, model design, and maintenance.
High operational complexity and cost for a task that can be done with simple code.
Scenario suitable: When you need predictive transformations or feature extraction, not simple geometric transformations.
Verdict: Rejected due to unnecessary complexity and cost.
---
Option B: Create an AWS Lambda function to perform the transformations
Pros:
Lambda can run custom code for image manipulation using libraries like Pillow or OpenCV.
Fully managed, serverless, operationally efficient, and scales automatically.
Can process images on-demand or in response to S3 events.
Cons:
Limited runtime (15 minutes max), but fine for most image transformations.
Scenario suitable: Ideal for event-driven image processing like rotation, resize, or transpose when images are uploaded to S3.
Verdict: Accepted ...
Author: Michael · Last updated Jul 19, 2026
An AI practitioner is writing software code. The AI practitioner wants to quickly develop a test case and create documentation for the code.
Whi...
Let’s carefully analyze the scenario:
Scenario:
AI practitioner is writing software code.
Needs quick test case development and documentation.
Goal: least effort solution on AWS.
---
Option A: Upload the code to an online coding assistant
Online coding assistants can suggest code, generate snippets, or even produce test cases and documentation.
Pros: Fast, minimal setup.
Cons: Might require using a non-AWS platform, which could complicate integration with AWS services.
Use Case: When quick code suggestions or automated documentation is needed without AWS-specific integration.
Verdict: Possible, but not fully aligned with AWS tools.
---
Option B: Develop an application to use foundation models (FMs)
FMs are large AI models for tasks like code generation or text summarization.
Pros: Can generate test cases and documentation.
Cons: Requires developing an application, managing model inference, integrating APIs.
Key Factor: High setup effort, not “least effort.”
Use Case: When you need custom AI-driven code solutions.
Verdict: Rejected because it requires more work than necessary.
---
Option C: Use Amazon Cod...
Author: Sophia · Last updated Jul 19, 2026
A company is developing a generative AI application to automatically generate product descriptions for an ecommerce website. The product descriptions must consist of paragraphs of text that are consistent in style and tone. The application must genera...
Let's carefully analyze the requirements and each option:
Requirements:
Automatically generate product descriptions.
Paragraphs of text with consistent style and tone.
Must scale to thousands of unique descriptions daily.
Context: AWS (so think about services like Amazon Bedrock or using large language models).
---
Option A: Variational Autoencoder (VAE)
Strengths: Good for generating continuous data like images, compressing features, and some structured data generation.
Weaknesses for this case: VAEs are not typically used for generating coherent, long-form text like paragraphs. They are better suited for images or simple structured outputs. Maintaining style and tone in text is challenging for VAEs.
Use case: Image generation, anomaly detection, or latent feature representation.
→ Not ideal for paragraph text generation.
---
Option B: Transformer-based model
Strengths: Excellent for natural language tasks. Can generate coherent, context-aware, and stylistically consistent text. Models like GPT, BERT (for fine-tuning), and other LLMs are transformer-based. Can scale to generate thousands of outputs efficiently.
Weaknesses: Require careful prompt design or fine-tuning to match brand tone.
Use case: Text generation, summarization, translation, conversational AI, or structured text like product descriptions.
→ Perfect match for generating long, coherent paragraphs with consistent st...
Author: Alexander · Last updated Jul 19, 2026
An AI practitioner has trained a model on a training dataset. The model performs well on the training data. However, the model does not perform well on...
Let’s analyze the scenario carefully.
Scenario:
The model performs well on the training dataset.
The model performs poorly on evaluation data (unseen data).
We are asked for the MOST likely cause.
---
Step 1: Analyze each option
A) The model is underfit
Underfitting happens when a model is too simple to capture patterns in the training data.
Key indicator: Poor performance on both training and evaluation data.
Our scenario: The model performs well on training data. ✅ So underfitting is unlikely.
B) The model requires prompt engineering
Prompt engineering applies mainly to large language models (LLMs) when they need better input phrasing.
Key indicator: The issue is due to how you ask the model, not its ability to generalize.
Our scenario: This is a trained model with clear evaluation data, not necessarily a generative prompt issue. ❌ So this is unlikely.
C) The model is biased
Bias refers to the model favoring certain outcomes due to biased training data.
Key indicator: Predictions may systematically favor some classes, regardless of training performance.
Our scenario: While bias can exist, the main symptom described is high training performance vs low evaluation p...
Author: MysticJaguar44 · Last updated Jul 19, 2026
A company wants to develop an interpretable ML model to assess the risk of loan applications.
Which type of ML...
Let’s carefully analyze the problem:
Goal: Build an interpretable ML model to assess the risk of loan applications. Interpretability is key here because financial institutions must explain decisions to regulators and customers.
Now, let’s evaluate each option:
---
A) Deep learning model
Pros: Can capture complex patterns and interactions in data.
Cons: Not very interpretable; neural networks are often “black boxes.”
Use case: Good for tasks like image recognition, NLP, or very complex prediction tasks where accuracy outweighs explainability.
Conclusion: Rejected because interpretability is required.
---
B) Logistic regression model
Pros:
Produces coefficients that indicate the impact of each feature on the probability of default.
Easy to explain to stakeholders and regulators.
Suitable for binary classification (approve/reject loan).
Cons: Cannot capture highly complex nonlinear patterns as effectively as deep learning.
Use case: Excellent for credit risk scoring, fraud detection, and any scenario where transparency ...
Author: Maya · Last updated Jul 19, 2026
A company stores customer personally identifiable information (PII) data. The company must store the PII data within the company's AWS...
Let’s carefully analyze this AWS-related question step by step.
Scenario: A company stores customer PII and must store it within the company's AWS Region.
We are asked: which aspect of governance this describes.
---
Option A: Data Mining
Definition: Data mining is the process of analyzing large datasets to extract patterns, trends, or knowledge.
Key factors: Data mining focuses on analysis and insights, not where the data is stored.
Why it’s rejected: The requirement is about location of storage, not analyzing or extracting information from the data.
Scenario where used: Companies use data mining for customer behavior analysis, fraud detection, or recommendation systems.
---
Option B: Data Residency
Definition: Data residency refers to where data is physically or logically stored, often to comply with regulatory, legal, or governance requirements.
Key factors:
Requirement: PII must remain within a specific AWS Region.
Data residency ensures compliance with laws (e.g., GDPR, HIPAA) about data staying in a geographic boundary.
Why it fits: Exactly matches the scenario of storing PII within a specific region.
Scenario where used: Any company handling sensitive data that must comply w...
Author: MysticJaguar44 · Last updated Jul 19, 2026
A company wants to implement a generative AI solution to improve its marketing operations. The company wants to increase its revenue in the n...
Let’s carefully analyze each option in the context of AWS generative AI, marketing operations, and the goal of increasing revenue within 6 months. I’ll break it down step by step.
---
Option A: Immediately start training a custom foundation model (FM) using the company's existing data
Pros: A custom FM could theoretically be tailored to the company’s marketing data.
Cons:
Training a custom foundation model from scratch or fine-tuning an FM is time-consuming and resource-intensive (weeks to months).
There is no guarantee it will immediately improve revenue; it’s a long-term solution.
Conclusion: Not suitable if the goal is measurable revenue increase within 6 months. It’s better for long-term differentiation scenarios or highly specialized AI requirements.
---
Option B: Conduct stakeholder interviews to refine use cases and set measurable goals
Pros:
Clarifies business objectives, aligns marketing goals with AI capabilities.
Helps identify high-impact use cases (e.g., targeted ad campaigns, personalized content) that could affect revenue.
Low cost, fast to implement.
Cons:
This is a planning step, not a direct AI solution, so by itself, it doesn’t generate revenue.
Conclusion: Important for planning, but won’t directly increase revenue within 6 months.
---
Option C: Implement a prebuilt AI assistant solution and measure its impact on customer satisfaction
Pros:
Prebuilt solutions (like AWS’s Amazon Bedrock models or AI marketing tools) can be quickly deployed.
Can automate marketing tasks li...
Author: Isabella1 · Last updated Jul 19, 2026
A healthcare company wants to create a model to improve disease diagnostics by analyzing patient voices. The company has recorded hundreds of patient voices for this project.
The company is currently filtering voice recordings accor...
Let's carefully analyze the scenario step by step.
Scenario:
A healthcare company has recorded hundreds of patient voices for disease diagnostics.
They are filtering voice recordings according to duration and language.
We are asked which phase of the ML lifecycle this corresponds to, considering AWS’s perspective.
---
Step 1: Analyze the options
A) Data collection
Data collection involves gathering raw data, such as recording patient voices in this case.
Key factor: The company already has the recordings, so data collection has mostly been completed.
❌ Rejected because the current task is not gathering data, it’s working with existing data.
---
B) Data preprocessing
Data preprocessing involves cleaning and preparing data for use in ML models.
Examples: filtering out poor-quality recordings, normalizing audio, removing irrelevant data, ensuring consistent formats.
Key factor: The company is filtering recordings based on duration and language, which is exactly cleaning and preparing data.
✅ This fits the scenario perfectly.
---
C) Feature engineering
Feature engineering is about extracting meaningful features from raw data, e.g., turning voice recordings into MFCCs, pitch, or other audio features that models can use.
Key factor: The company is ...
