Coffee Talk: Java, News, Stories and Opinions

BLOG

AWS AI Practitioner Certification Exam Dumps and Braindumps

TechTarget

AWS AI Practitioner Exam Simulator

Despite the title of this article, this isn’t an AWS AI Practitioner exam braindump in the traditional sense.

I don’t believe in cheating.

A true braindump is when someone takes the actual exam and tries to rewrite every question they remember, essentially dumping the test content online. That’s unethical and a clear violation of AWS’s exam policies. There’s no integrity or value in that approach.

This set of AWS AI Practitioner exam questions is nothing like that.

Much better than a certification exam dump!

All of the questions here come from my AWS AI Practitioner Udemy course and from my AWS certification site: certificationexams.pro. It hosts hundreds of original practice questions designed around AWS certification objectives and AWS AI Practitioner exam topics, so be sure to check it out if you want to be thoroughly prepared for the exam.

My AWS AI Practitioner Udemy Course has over 500 practice questions.

But as I was saying, this is not a true ‘exam dump.’ Each AI Practitioner certification question here has been thoughtfully written to align with the official exam guide, testing your understanding of AI and ML fundamentals, AWS AI services, responsible AI concepts, and practical business use cases, without ever copying or disclosing real AWS exam content.

The goal is to help you learn ethically, build real knowledge, and feel confident working with AWS AI tools like Amazon SageMaker, AWS Bedrock, Comprehend, Rekognition, and Lex.

If you can answer these questions with confidence, and understand why each option is right or wrong, you won’t just pass the AWS AI Practitioner exam. You’ll gain a true understanding of how certified AI solutions are designed, deployed, and managed on AWS.

So, call it a braindump if you like, but it’s really a smart, honest study companion created to help you think like a cloud AI professional.

These AWS AI Practitioner exam questions are designed to challenge you, but each one includes clear explanations, practical insights, and exam tips to help you succeed on test day.

Learn deeply, study smart, and good luck on your AWS AI Practitioner certification journey.

Git, GitHub & GitHub Copilot Certification Made Easy
Want to get certified on the most popular AI, ML & DevOps technologies of the day? These five resources will help you get GitHub certified in a hurry. Master vibe coding and prove it with the GitHub Copilot certification Prove you know how CI/CD with the GitHub Action certification Prove you know your way around Git and GitHub with the GitHub Foundation certification. Show you know how to work on a team with a Scrum Master cert or Product Owner credential Let employers know you understand LLMs and Agentic AI with a GCP AI Leader certification Get certified in the latest AI, ML and DevOps technologies. Advance your career today.

Git, GitHub & GitHub Copilot Certification Made Easy

Want to get certified on the most popular AI, ML & DevOps technologies of the day? These five resources will help you get GitHub certified in a hurry.

Master vibe coding and prove it with the GitHub Copilot certification
Prove you know how CI/CD with the GitHub Action certification
Prove you know your way around Git and GitHub with the GitHub Foundation certification.
Show you know how to work on a team with a Scrum Master cert or Product Owner credential
Let employers know you understand LLMs and Agentic AI with a GCP AI Leader certification

Get certified in the latest AI, ML and DevOps technologies. Advance your career today.

AWS AI Practitioner Exam Questions

Question 1

A retail startup has fine-tuned a foundation model in Amazon Bedrock to handle customer support product questions and now wants to run validation on a new set of 5,000 test prompts. Which AWS service should they use to upload and store the validation dataset so it can be used by the evaluation workflow?

[ ] A. AWS Glue
[ ] B. Amazon S3
[ ] C. Amazon Elastic Block Store (Amazon EBS)
[ ] D. Amazon Elastic File System (Amazon EFS)

Question 2

Scrumtuous Agile Services Inc is developing a computer vision system to identify crushed or dented boxes on conveyor lines at 14 distribution hubs. The team needs highly reliable annotations so the training data minimizes both false positives and false negatives. What is the most effective way to obtain precise image labels and reduce labeling errors?

[ ] A. Accept labels from a general crowdsourcing vendor without strict quality checks
[ ] B. Use Amazon SageMaker Ground Truth with only automated labeling and no review
[ ] C. Set up a human-in-the-loop review using Amazon Augmented AI (Amazon A2I) to validate annotations
[ ] D. Amazon Rekognition auto-labeling without any human verification

Question 3

A regional lender named McKenzie Software Credit launches an AI assistant to summarize 120-page mortgage agreements and flag key clauses. The team notices that responses vary between runs on the same document. Which preprocessing step would most improve result consistency?

[ ] A. Token normalization
[ ] B. Context expansion
[ ] C. Document chunking with overlap
[ ] D. Fine-tuning

Question 4

Solstice Capital, a fintech firm building AI workloads on AWS, is formalizing data governance controls. What is the key distinction between data residency and data logging in this scenario?

[ ] A. Data residency focuses on encryption and key management requirements, while data logging deals with ETL and data transformation steps
[ ] B. Data residency dictates the country or region where datasets must be stored, whereas data logging records who accessed the data and what changed over time
[ ] C. Data residency tracks user activity within an AI application, and data logging determines the permitted geographic locations for storage
[ ] D. Data residency is about monitoring live data usage patterns, while data logging is used to enforce retention with Amazon S3 Lifecycle rules

Question 5

Lakeside Insurance Group plans to release a customer support chatbot powered by a fine-tuned Amazon SageMaker JumpStart foundation model. The rollout must satisfy compliance obligations across three jurisdictions, with emphasis on privacy and security controls. For which capabilities can the team credibly demonstrate compliance as part of the application’s design? (Choose 2)

[ ] A. Decoupled microservices architecture
[ ] B. Threat detection
[ ] C. Automatic scaling for inference endpoints
[ ] D. Data protection
[ ] E. Cost efficiency

Question 6

Which AWS service provides secure, durable, virtually unlimited object storage for about 25 TB of text data that Amazon Bedrock can access for model customization?

[ ] A. Amazon RDS
[ ] B. Amazon EFS
[ ] C. Amazon Simple Storage Service (S3)
[ ] D. Amazon Redshift

Question 7

At Northwind Bikes, a team trains a model on Amazon SageMaker and reduces the training set from 1.2 million to 350,000 records while also adding many new features with minimal regularization. They want to know if such choices could cause both bias and variance to rise at the same time and what that would mean for model quality?

[ ] A. No, bias and variance always move in opposite directions and cannot both rise at the same time
[ ] B. Yes, raising both bias and variance generally boosts accuracy and generalization
[ ] C. Yes, both bias and variance can increase simultaneously, which usually harms performance by combining underfitting and overfitting
[ ] D. Amazon SageMaker Clarify

Question 8

A regional logistics provider, Alpine Parcel, plans to pilot a generative AI assistant across three departments within 90 days to answer customer questions and draft catalog content. What challenges should the team expect and plan mitigations for when applying generative AI to these business workflows? (Choose 2)

[ ] A. Using Amazon Bedrock removes the need for governance and compliance controls
[ ] B. Outputs can include fabricated or incorrect information
[ ] C. Generative AI never requires labeled data or further training for a specific domain
[ ] D. Meeting regulatory and privacy obligations can be difficult
[ ] E. Generative AI models always provide clear, built-in explanations for their responses

Question 9

A digital payments startup named Pickering is Springfield Inc processes about 20 million card transactions per day and uses a foundation model to flag suspicious activity. The team wants to assess the model’s precision and recall to validate performance. Which situation best demonstrates why these metrics matter?

[ ] A. Reducing end-to-end model inference latency for transactions
[ ] B. Lowering token usage charges for prompts in Amazon Bedrock
[ ] C. Managing false alarms and missed fraud cases in the alerting system
[ ] D. Evaluating BLEU scores for a neural machine translation model

Question 10

A digital publisher on AWS uses a generative model to draft history articles, and editors reviewing 240 recent summaries have found misdated timelines and misattributed events. Which approach will most reliably improve factual correctness in the model’s responses?

[ ] A. Perform fine-tuning on the model
[ ] B. Reduce the temperature setting
[ ] C. Use Retrieval Augmented Generation (RAG)
[ ] D. Amazon SageMaker Clarify

Check out my AWS AI Practitioner Udemy course for more questions.

Question 11

A subscription-based streaming platform, Scrumtuous Flashcards & Media, wants to segment viewers and detect new content trends from about 45 TB of clickstream and watch data it stores in Amazon S3. The dataset has no existing labels for audience personas, and the analytics team plans to experiment with clustering in Amazon SageMaker to reveal natural groups. Which characteristics of unsupervised learning should guide their approach? (Choose 2)

[ ] A. It builds models by maximizing classification accuracy
[ ] B. It is well suited when examples are unlabeled or labeling is costly
[ ] C. It focuses on reward-driven decision making over time
[ ] D. It discovers latent patterns or relationships without predefined targets
[ ] E. It depends on labeled historical outcomes to forecast future behavior

Question 12

Which SageMaker inference option is best for scheduled batch processing of large (25 GB) objects from S3 without persistent endpoints?

[ ] A. SageMaker Serverless Inference
[ ] B. Amazon SageMaker Batch Transform
[ ] C. SageMaker Processing
[ ] D. SageMaker Asynchronous Inference

Question 13

A metropolitan transit agency wants to improve safety by spotting and tracking unattended bags on train platforms from live CCTV feeds. They need to automatically flag when a backpack or suitcase remains alone for more than 4 minutes without a nearby person and send immediate alerts to their control room. The team plans to build this with Amazon Rekognition. Which Rekognition features are most appropriate for this requirement? (Choose 2)

[ ] A. Text Detection
[ ] B. Facial Recognition
[ ] C. Activity Stream Analysis
[ ] D. Object and Scene Detection
[ ] E. Audio Analysis

Question 14

A regional retail bank named Blue Harbor Bank is launching a generative AI application on Amazon Bedrock to deliver tailored portfolio guidance to customers. The bank needs near real-time visibility into model latency and failure rates, overall system health, and the ability to search and analyze logs as issues occur. They want centralized observability with alarms that trigger quickly, metrics dashboards, and log analytics for the AI pipelines. Which AWS service should they integrate to provide comprehensive monitoring, logging, and observability?

[ ] A. AWS Config
[ ] B. Amazon CloudWatch
[ ] C. AWS CloudTrail
[ ] D. Amazon Inspector

Question 15

A data lead at Helios Logistics needs to pair three AWS AI services with their primary capabilities. Match A) Amazon Textract, B) Amazon Forecast, and C) Amazon Kendra to 1) an enterprise search service powered by machine learning across company content, 2) automated extraction of printed text, handwriting, and form or table data from documents, and 3) machine learning based forecasting for business time series. Which mapping is correct?

[ ] A. A→1, B→3, C→2
[ ] B. A→3, B→1, C→2
[ ] C. A→2, B→3, C→1
[ ] D. A→1, B→2, C→3

Question 16

McKenzie Mission Control, a digital media startup, is piloting generative AI for ad copy and automated chat replies. Over the next 60 days, the team will test 250 prompts across several models and needs to distinguish the step that produces responses from the process that scores quality to pick a production model. Which statement best describes the difference between model inference and model evaluation?

[ ] A. Both model inference and model evaluation are the act of generating a response from a prompt
[ ] B. Amazon Bedrock
[ ] C. Model inference is when a model generates an output from an input prompt, whereas model evaluation compares model outputs against defined metrics to choose the most suitable model
[ ] D. Both model inference and model evaluation are used to compare model outputs and decide which model is best for a use case

Question 17

A multinational online marketplace called Scrumtuous Mart plans to pilot generative AI to deliver product suggestions in under 150 milliseconds, power multilingual customer chat, and translate shopper reviews as they are posted. Which considerations should most influence the choice of a generative AI model? (Choose 2)

[ ] A. The number of provisioned CPU cores
[ ] B. Governance, safety, and regulatory compliance obligations
[ ] C. Availability of popular open-source ML frameworks
[ ] D. Latency and quality targets, including response time and accuracy
[ ] E. Amazon CloudFront

Question 18

In SageMaker Automatic Model Tuning with autotune enabled and the objective metric defined, what additional configuration must be explicitly set for the tuning job to run?

[ ] A. Early stopping settings
[ ] B. No extra setting required
[ ] C. Hyperparameter ranges
[ ] D. Number of training jobs

Question 19

A metropolitan transit authority uses an AI video analytics pipeline on AWS to detect unsafe behavior across 180 station cameras. Community groups report that riders from one ethnic background are being flagged far more often than others. The data team believes the training dataset built in Amazon SageMaker did not represent all demographics evenly, which skewed the model’s outputs. Which form of bias best explains this outcome?

[ ] A. Confirmation bias
[ ] B. Amazon Rekognition
[ ] C. Measurement bias
[ ] D. Sampling bias in the training data

Question 20

Orion Analytics, a retail insights startup, uses a foundation model to summarize customer interactions and stores the outputs and intermediate datasets in Amazon S3 for 90 days. The security team must protect personally identifiable information at rest with centralized key management that integrates across AWS services. Which AWS service should they use?

[ ] A. SageMaker Clarify
[ ] B. AWS KMS
[ ] C. AWS Secrets Manager
[ ] D. AWS CloudTrail

Answers to the AWS AI Practitioner Exam Questions

Question 1

[*] B. Amazon S3

Amazon S3 is the right choice because Amazon Bedrock workflows expect datasets to be accessible from object storage, and S3 provides S3 URIs that Bedrock can reference directly for evaluation and fine-tuning.

AWS Glue is for ETL jobs, data cataloging, and pipelines; it is not a storage service for directly hosting evaluation datasets for Bedrock.

Amazon Elastic Block Store (Amazon EBS) is block storage tied to individual EC2 instances and does not provide the object-based access or S3 URIs required by Bedrock evaluation jobs.

Amazon Elastic File System (Amazon EFS) is a POSIX file system for shared access, but Bedrock does not natively read datasets from EFS; it expects data in S3.

Exam Tip

For questions involving dataset upload or access by foundation model customization or evaluation in Bedrock, look for Amazon S3 as the storage answer and watch for distractors that are compute-attached storage, file systems, or ETL services.

Question 2

Orion Parcel Services is developing a computer vision system to identify crushed or dented boxes on conveyor lines at 14 distribution hubs. The team needs highly reliable annotations so the training data minimizes both false positives and false negatives. What is the most effective way to obtain precise image labels and reduce labeling errors?

[*] C. Set up a human-in-the-loop review using Amazon Augmented AI (Amazon A2I) to validate annotations

The best approach is Set up a human-in-the-loop review using Amazon Augmented AI (Amazon A2I) to validate annotations. A2I lets you route uncertain or high-impact cases to human reviewers, improving label precision for subtle damage classes and reducing both false positives and false negatives.

Accept labels from a general crowdsourcing vendor without strict quality checks is risky because missing QA gates, gold questions, or consensus checks commonly introduces inconsistent labels and noise.

Use Amazon SageMaker Ground Truth with only automated labeling and no review omits human validation, so any systematic model errors in auto-labeling can spread and lower overall data quality.

Amazon Rekognition auto-labeling without any human verification relies entirely on automation, which is insufficient for nuanced defect detection where manual confirmation materially boosts accuracy.

Exam Tip

When you see requirements for high label accuracy or safety-critical outcomes, look for a human-in-the-loop solution; Amazon A2I integrates human review directly into ML workflows.

Question 3

A regional lender named NorthRiver Credit launches an AI assistant to summarize 120-page mortgage agreements and flag key clauses. The team notices that responses vary between runs on the same document. Which preprocessing step would most improve result consistency?

[*] C. Document chunking with overlap

Document chunking with overlap is the most effective preprocessing step because it breaks lengthy agreements into coherent sections that fit within the model’s context window while preserving continuity across chunk boundaries. This reduces attention drift, avoids context overflow, and yields more repeatable results across runs.

Token normalization can clean text but does not ensure clause-level boundaries or manage context length, so inconsistencies from processing very large documents remain.

Context expansion introduces additional surrounding text that may add noise or exceed the context window, which often makes outputs less stable.

Fine-tuning alters model parameters and is not a preprocessing step; it is costly and unnecessary for issues primarily caused by input segmentation and context handling.

Exam Tip

When responses vary on long documents, prioritize chunk size, overlap, and clear boundaries before considering heavier changes like fine-tuning.

Question 4

Solstice Capital, a fintech firm building AI workloads on AWS, is formalizing data governance controls. What is the key distinction between data residency and data logging in this scenario?

[*] B. Data residency dictates the country or region where datasets must be stored, whereas data logging records who accessed the data and what changed over time

The correct choice is Data residency dictates the country or region where datasets must be stored, whereas data logging records who accessed the data and what changed over time. Data residency is about complying with geographic storage requirements, while logging creates an audit trail for access and modifications.

Data residency focuses on encryption and key management requirements, while data logging deals with ETL and data transformation steps is incorrect because encryption and transformation concern security and processing, not storage location or audit events.

Data residency tracks user activity within an AI application, and data logging determines the permitted geographic locations for storage is incorrect because it flips the definitions; tracking access is logging, and location constraints are residency.

Data residency is about monitoring live data usage patterns, while data logging is used to enforce retention with Amazon S3 Lifecycle rules is incorrect because residency is not real-time monitoring, and logging does not enforce lifecycle policies.

Exam Tip

Remember: data residency answers where data is stored, while data logging answers who, what, and when for access and changes.

Question 5

[*] B. Threat detection
[*] D. Data protection

The correct capabilities are Threat detection and Data protection. Compliance frameworks emphasize protecting sensitive data through encryption, key management, access controls, and secure handling, as well as implementing detective controls to identify and respond to suspicious activity.

Decoupled microservices architecture improves maintainability and resilience but is not a regulatory control on its own.

Automatic scaling for inference endpoints addresses performance and reliability rather than meeting explicit compliance requirements.

Cost efficiency is a financial and operational concern, not a compliance control.

Exam Tip

When a question mentions compliance across regions or frameworks, look for capabilities tied to security controls like data protection and threat detection rather than architectural patterns or cost objectives.

Question 6

Which AWS service provides secure, durable, virtually unlimited object storage for about 25 TB of text data that Amazon Bedrock can access for model customization?

[*] C. Amazon Simple Storage Service (S3)

Amazon Simple Storage Service (S3) is correct because Amazon Bedrock model customization expects training and validation data in S3. S3 provides virtually unlimited scalability, eleven nines of durability, strong security controls, and easy integration with Bedrock customization jobs.

The option Amazon RDS is wrong because it is a relational database for structured transactional workloads, not large object storage for training data.

Amazon EFS is wrong because it is a network file system for POSIX workloads rather than object storage and is not the typical source for Bedrock training data.

Amazon Redshift is wrong because it is a columnar data warehouse for analytics, not for storing large unstructured training files.

Exam Tip

When you see cues like secure, durable, and virtually unlimited object storage for ML training, choose S3. Also, if the question mentions Bedrock model customization or fine-tuning data location, map it directly to S3.

Question 7

[*] C. Yes, both bias and variance can increase simultaneously, which usually harms performance by combining underfitting and overfitting

The correct choice is Yes, both bias and variance can increase simultaneously, which usually harms performance by combining underfitting and overfitting. Under certain data conditions and hyperparameter settings, a model can become too simplistic in its assumptions (high bias) while also being overly sensitive to fluctuations in the available data (high variance), producing poor generalization.

No, bias and variance always move in opposite directions and cannot both rise at the same time is incorrect because the bias–variance trade-off is not a hard constraint; both can be high in poorly designed setups or with limited data.

Yes, raising both bias and variance generally boosts accuracy and generalization is wrong since increasing both typically worsens performance by amplifying underfitting and overfitting effects.

Amazon SageMaker Clarify is not a valid answer to this conceptual question; while it helps analyze bias and explain model predictions, it does not change the fact that both bias and variance can increase simultaneously.

Exam Tip

Watch for absolute language like always or cannot. In the bias–variance topic, both can be high at once, and the best answer usually mentions degraded performance due to simultaneous underfitting and overfitting.

Question 8

[*] B. Outputs can include fabricated or incorrect information
[*] D. Meeting regulatory and privacy obligations can be difficult

Outputs can include fabricated or incorrect information is correct because LLMs may hallucinate, returning convincing yet wrong or nonsensical content. Businesses should use human review, retrieval-augmented generation, and guardrails to reduce this risk.

Meeting regulatory and privacy obligations can be difficult is correct since organizations must handle data residency, privacy, and sector-specific rules, which requires governance, auditing, and content filtering.

Using Amazon Bedrock removes the need for governance and compliance controls is wrong because while Bedrock provides capabilities like guardrails and model access, customers still must implement policies and controls under the shared responsibility model.

Generative AI never requires labeled data or further training for a specific domain is wrong since many enterprise scenarios improve with domain adaptation, fine-tuning, or labeled datasets, even when starting from foundation models.

Generative AI models always provide clear, built-in explanations for their responses is wrong because LLMs are not inherently interpretable, and explainability typically requires additional techniques and tooling.

Exam Tip

For generative AI questions, watch for hallucinations and compliance as classic risks; claims with words like always or never are often distractors.

Question 9

A digital payments startup processes about 20 million card transactions per day and uses a foundation model to flag suspicious activity. The team wants to assess the model’s precision and recall to validate performance. Which situation best demonstrates why these metrics matter?

[*] C. Managing false alarms and missed fraud cases in the alerting system

The correct choice is Managing false alarms and missed fraud cases in the alerting system. Precision measures how many flagged transactions are truly fraudulent, while recall measures how many fraudulent transactions are actually caught. In fraud detection, you must balance avoiding false alarms with not letting real fraud slip through, which is exactly what precision and recall quantify.

Reducing end-to-end model inference latency for transactions focuses on responsiveness and throughput, not on the correctness trade-offs between false positives and false negatives. It does not explain why precision and recall matter.

Lowering token usage charges for prompts in Amazon Bedrock is a cost optimization concern and does not evaluate the model’s ability to accurately distinguish fraudulent from legitimate transactions.

Evaluating BLEU scores for a neural machine translation model applies to translation quality, not binary or multi-class classification performance; it is unrelated to fraud detection precision and recall.

Exam Tip

When you see precision and recall, think about the trade-off between false positives and false negatives; scenarios that emphasize avoiding false alarms and catching as many true events as possible usually point to these metrics.

Question 10

[*] C. Use Retrieval Augmented Generation (RAG)

Use Retrieval Augmented Generation (RAG) is correct because it augments prompts with trusted, up-to-date context retrieved from vetted data stores, which anchors the model’s answers to verifiable sources and reduces hallucinations about dates and events. On AWS, this can be implemented with Knowledge Bases for Amazon Bedrock that connect to sources like Amazon S3 and vector indexes to ground responses.

Perform fine-tuning on the model is not ideal here because fine-tuning adapts behavior and style but does not guarantee correctness of specific facts and can quickly become outdated without continual data refresh.

Reduce the temperature setting lowers randomness but does not improve factual grounding; the model can still produce confidently wrong statements, just more consistently.

Amazon SageMaker Clarify addresses bias and explainability concerns for ML workflows and does not provide retrieval or grounding needed to correct historical inaccuracies.

Exam Tip

When the key requirement is factual accuracy tied to external knowledge, think RAG; use fine-tuning for task/style adaptation and temperature only to control variability, not truthfulness.

Question 11

A subscription-based streaming platform, PolarWave Media, wants to segment viewers and detect new content trends from about 45 TB of clickstream and watch data it stores in Amazon S3. The dataset has no existing labels for audience personas, and the analytics team plans to experiment with clustering in Amazon SageMaker to reveal natural groups. Which characteristics of unsupervised learning should guide their approach? (Choose 2)

[*] B. It is well suited when examples are unlabeled or labeling is costly
[*] D. It discovers latent patterns or relationships without predefined targets

It is well suited when examples are unlabeled or labeling is costly is correct because unsupervised learning explicitly addresses situations where no ground-truth labels exist or where creating them is prohibitively expensive, such as customer or viewer segmentation from raw event logs.

It discovers latent patterns or relationships without predefined targets is also correct because unsupervised techniques like clustering and dimensionality reduction reveal natural groupings and structure without target variables, which aligns with the goal of audience segmentation and trend discovery.

It builds models by maximizing classification accuracy is incorrect because accuracy is a supervised metric requiring known labels, which are absent in this scenario.

It focuses on reward-driven decision making over time is incorrect because that describes reinforcement learning, which learns via rewards and actions rather than exploring unlabeled datasets for structure.

It depends on labeled historical outcomes to forecast future behavior is incorrect since this is the essence of supervised learning and does not apply when labels are not available.

Exam Tip

When you see no labels and goals like segmentation or pattern discovery, think unsupervised learning; when you see known targets and metrics like accuracy, think supervised; when you see rewards and policies, think reinforcement learning.

Question 12

Which SageMaker inference option is best for scheduled batch processing of large (25 GB) objects from S3 without persistent endpoints?

[*] B. Amazon SageMaker Batch Transform

The correct choice is Amazon SageMaker Batch Transform. It is designed for offline, scheduled batch inference directly against data in Amazon S3, scales with the job, and does not require maintaining always-on endpoints. This makes it cost-effective and well-suited for large payloads and periodic processing.

The option SageMaker Serverless Inference focuses on real-time request-response with spiky traffic patterns and is not ideal for sweeping through large datasets on a schedule.

SageMaker Processing is intended for data preprocessing, postprocessing, and evaluation—not for running large-scale inference jobs.

SageMaker Asynchronous Inference accommodates large payloads and long-running requests but still uses endpoints and is oriented toward per-request workloads rather than bulk, scheduled batch jobs.

Exam Tip

Watch for cues like batch, scheduled, large files, and no always-on endpoints—these point to Batch Transform. Choose real-time or serverless endpoints for low-latency request-response. Pick asynchronous inference when you need endpoint-based, per-request handling of large or long-running jobs without immediate responses.

Question 13

[*] B. Facial Recognition
[*] D. Object and Scene Detection

Object and Scene Detection can label items such as bags, suitcases, and people in image frames or video streams, which is essential for noticing when luggage is present and whether a person is nearby.

Facial Recognition can detect and match faces, enabling workflows that associate a bag with its owner and determine if the person leaves, which helps infer that an item has been left unattended.

Text Detection focuses on recognizing written characters and does not help determine if luggage is unattended.

Activity Stream Analysis is not a Rekognition capability and generally pertains to analyzing event or user activity logs rather than computer vision.

Audio Analysis deals with sound, which Rekognition does not process, making it irrelevant for video-only bag detection.

Exam Tip

For video surveillance use cases, think in terms of Rekognition labels for objects and scenes and Rekognition faces for associating people with objects; ignore features that are not part of Rekognition or that deal with audio or text when the task is visual-only.

Question 14

[*] B. Amazon CloudWatch

The correct choice is Amazon CloudWatch because it provides unified observability across metrics, logs, and traces, supports near real-time alarms, and offers dashboards and log query capabilities that fit AI workloads on Amazon Bedrock.

AWS Config focuses on configuration state and compliance rather than performance metrics, alarms, or log analysis for running applications.

AWS CloudTrail captures API events for auditing and governance but does not provide the operational metrics, dashboards, or alerting needed for workload health.

Amazon Inspector is a vulnerability management tool for security findings and does not deliver monitoring dashboards, alarms, or log analytics.

Exam Tip

Associate metrics, alarms, dashboards, and log analytics with Amazon CloudWatch; map API auditing to CloudTrail, configuration compliance to Config, and vulnerability scanning to Inspector.

Question 15

[*] C. A→2, B→3, C→1

The correct mapping is A→2, B→3, C→1 because Amazon Textract extracts text and structured data from documents, Amazon Forecast generates machine learning based forecasts for business outcomes, and Amazon Kendra provides accurate enterprise search across organizational content.

A→1, B→3, C→2 is wrong because it swaps Textract and Kendra while only Forecast is placed correctly.

A→3, B→1, C→2 is wrong because all three services are mismatched with capabilities they do not provide.

A→1, B→2, C→3 is wrong because it assigns Kendra to Textract’s OCR role and gives Forecast a document extraction task.

Exam Tip

Memorize the core one-liners: Textract for document text and form extraction, Forecast for ML-driven time series predictions, and Kendra for enterprise search.

Question 16

Kestrel Dynamics, a digital media startup, is piloting generative AI for ad copy and automated chat replies. Over the next 60 days, the team will test 250 prompts across several models and needs to distinguish the step that produces responses from the process that scores quality to pick a production model. Which statement best describes the difference between model inference and model evaluation?

[*] C. Model inference is when a model generates an output from an input prompt, whereas model evaluation compares model outputs against defined metrics to choose the most suitable model

The correct choice is Model inference is when a model generates an output from an input prompt, whereas model evaluation compares model outputs against defined metrics to choose the most suitable model. Inference is the act of running the model to produce a response. Evaluation is the process of scoring or comparing those responses using metrics to select or validate a model for a given use case.

The option Both model inference and model evaluation are the act of generating a response from a prompt is incorrect because evaluation is not generation; it is assessment against criteria.

The option Amazon Bedrock is incorrect because it names a service rather than explaining the conceptual difference between inference and evaluation.

The option Both model inference and model evaluation are used to compare model outputs and decide which model is best for a use case is incorrect because inference is about producing outputs, not comparing them.

Exam Tip

Link the verbs to the terms: inference = generate outputs from a prompt; evaluation = measure outputs against metrics to compare models.

Question 17

A multinational online marketplace called LumaMart plans to pilot generative AI to deliver product suggestions in under 150 milliseconds, power multilingual customer chat, and translate shopper reviews as they are posted. Which considerations should most influence the choice of a generative AI model? (Choose 2)

[*] B. Governance, safety, and regulatory compliance obligations
[*] D. Latency and quality targets, including response time and accuracy

Governance, safety, and regulatory compliance obligations is critical because operating in multiple regions demands controls for data privacy, content safety, and policy enforcement. Selecting a model that supports guardrails, moderation, and compliance workflows helps satisfy legal and organizational requirements.

Latency and quality targets, including response time and accuracy directly map to business SLAs for real-time recommendations, chat, and translation. The chosen model must meet low-latency expectations while maintaining accuracy relevant to the use case and language coverage.

The number of provisioned CPU cores is not a primary driver for generative model selection, as inference typically relies more on GPUs or specialized accelerators and model architecture choices.

Availability of popular open-source ML frameworks influences developer experience but does not determine the most suitable foundation model for production outcomes and risk requirements.

Amazon CloudFront is a CDN unrelated to choosing a generative model; it assists delivery at the edge but does not impact model capabilities, safety, or accuracy.

Exam Tip

When choosing a generative model, focus first on use-case performance SLAs and governance/compliance; tools, frameworks, and infrastructure knobs are secondary to meeting latency, accuracy, and risk requirements.

Question 18

In SageMaker Automatic Model Tuning with autotune enabled and the objective metric defined, what additional configuration must be explicitly set for the tuning job to run?

[*] B. No extra setting required

The correct answer is No extra setting required. When autotune is enabled and the objective metric is defined, SageMaker can infer key defaults such as search strategy, hyperparameter ranges, and the number of training jobs, allowing the tuning job to start without additional required settings.

The option Early stopping settings is incorrect because early stopping is optional and does not need to be specified to begin tuning.

The option Hyperparameter ranges is incorrect because autotune can automatically suggest reasonable ranges.

The option Number of training jobs is incorrect because autotune can determine a suitable number of evaluations to meet the objective.

Exam Tip

When you see autotune enabled and the objective metric provided, assume many tuning parameters can be auto-configured. Watch for distractors like ranges or job counts that are mandatory in manual tuning but optional with autotune.

Question 19

[*] D. Sampling bias in the training data

The correct answer is Sampling bias in the training data. When the training dataset does not reflect the real-world population and some demographics are overrepresented or underrepresented, the model can learn patterns that unfairly target those groups, causing disproportionate flags.

Confirmation bias is about human interpreters seeking evidence that fits prior beliefs, which is not the root cause when the model is behaving unfairly due to imbalanced training data.

Amazon Rekognition is an AWS computer vision service, not a bias category; the issue described is driven by dataset composition rather than the choice of service itself.

Measurement bias arises from systematic errors in data collection or labeling, such as faulty sensors or inconsistent annotations, which is different from unequal demographic representation in the dataset.

Exam Tip

When you see phrases like underrepresented groups in training data or unbalanced dataset, think sampling bias. If you see faulty sensors or mislabeled data, that cues measurement bias; if you see interpreting evidence to fit beliefs, that points to confirmation bias.

Question 20

[*] B. AWS KMS

The correct choice is AWS KMS because it provides centralized creation, management, and access control of encryption keys and integrates with services such as Amazon S3, Amazon EBS, and Amazon RDS to protect data at rest.

SageMaker Clarify focuses on bias detection and model explainability and does not handle encryption of stored datasets.

AWS Secrets Manager is for storing and rotating secrets like credentials and tokens; while it uses KMS under the hood, it is not the primary service used to encrypt large data stores at rest.

AWS CloudTrail delivers audit logs for API calls and governance, but it does not provide encryption for data at rest.

Exam Tip

When you see requirements for protecting sensitive data at rest with centralized key management across multiple AWS services, map that to AWS KMS; if the question centers on storing or rotating credentials, think AWS Secrets Manager instead.

Jira, Scrum & AI Certification
Want to get certified on the most popular software development technologies of the day? These resources will help you get Jira certified, Scrum certified and even AI Practitioner certified so your resume really stands out.. Show you know how to work on a team with a Scrum Master cert or Product Owner credential Get Atlassian certified as a Jira Cloud Expert or a as a Jira Project Manager Let employers know you understand LLMs and Agentic AI with a GCP AI Leader certification Maybe even master vibe coding and prove it with the GitHub Copilot certification And show off how well you know the cloud with these AWS, GCP and Azure certifications You can even get certified in the latest AI, ML and DevOps technologies. Advance your career today.

Jira, Scrum & AI Certification

Want to get certified on the most popular software development technologies of the day? These resources will help you get Jira certified, Scrum certified and even AI Practitioner certified so your resume really stands out..

Show you know how to work on a team with a Scrum Master cert or Product Owner credential
Get Atlassian certified as a Jira Cloud Expert or a as a Jira Project Manager
Let employers know you understand LLMs and Agentic AI with a GCP AI Leader certification
Maybe even master vibe coding and prove it with the GitHub Copilot certification
And show off how well you know the cloud with these AWS, GCP and Azure certifications

You can even get certified in the latest AI, ML and DevOps technologies. Advance your career today.

Cameron McKenzie is an AWS Certified AI Practitioner, Machine Learning Engineer, Solutions Architect and author of many popular books in the software development and Cloud Computing space. His growing YouTube channel training devs in Java, Spring, AI and ML has well over 30,000 subscribers.