13 Analytics, AI, and Machine Learning · AWS Cloud Practitioner

In this chapter you will learn how AWS analytics services collect, store, catalog, query, and visualize data, from batch data lakes built on Amazon S3, AWS Glue, and Amazon Athena to real-time streams with Amazon Kinesis. You will then learn the difference between artificial intelligence, machine learning, and generative AI; how Amazon Bedrock, Amazon Q, and Amazon SageMaker AI fit together; which pre-trained AI service matches which type of input; and the basics of responsible AI. This chapter maps to exam Domain 3, Cloud Technology and Services, Task Statement 3.7: Identify AWS artificial intelligence and machine learning (AI/ML) services and analytics services.

13.1 From data to decisions

Organizations collect enormous amounts of data: website clicks, sales transactions, sensor readings, support tickets, scanned forms, recorded calls. On its own, that data is just cost. It becomes valuable when someone can ask it questions (“Which products sell best in which cities?”) or when software can learn from it (“Is this transaction likely to be fraud?”). This chapter covers two families of services that turn data into value. Analytics services help people query, transform, and visualize data. AI and machine learning services help software recognize patterns, make predictions, and generate new content. For the exam, you need to recognize each service by its job, not configure it.

13.2 The analytics building blocks

Most analytics designs on AWS follow the same five stages: ingest data, store it, catalog and prepare it, query or process it, and visualize it. Keep that sequence in mind; it makes the services easy to place.

Amazon S3 as the data lake

A data lake is a central repository that stores data of any type and size, structured or not, in its original format, so it can be analyzed later in many ways. On AWS, the data lake almost always lives in Amazon S3, because S3 is durable, scales without limits you need to plan for, and is inexpensive, especially with lower-cost storage classes for older data. Teams often organize the lake into zones, such as a raw zone for data as it arrived and a curated zone for cleaned data. The key architectural idea is separating storage from compute. Data sits once in S3; many different engines (Athena, Redshift, EMR, SageMaker) can analyze it, each scaling and billing independently.

AWS Glue

AWS Glue is a serverless data integration service. It has three parts worth knowing:

Crawlers scan data in S3 and other sources and infer its structure (columns and data types).

The AWS Glue Data Catalog stores that structure as table definitions, a central metadata catalog that Athena, Redshift, EMR, and other services all use to understand what is in the lake.

ETL jobs (extract, transform, load) clean, reshape, and convert data, for example turning raw CSV logs into compressed, columnar Parquet files that are faster and cheaper to query. Glue Studio offers a visual editor, and AWS Glue DataBrew lets analysts clean data with point-and-click recipes and no code.

Amazon Athena

Amazon Athena is a serverless, interactive query service that runs standard SQL directly against data in Amazon S3. There is nothing to load and no infrastructure to manage: you point Athena at a table in the Glue Data Catalog and start querying. You pay for the amount of data each query scans, so compressing and partitioning data, and using columnar formats, directly lowers cost. Athena is perfect for ad hoc questions, such as analyzing load balancer logs or CloudTrail logs stored in S3.

Amazon Redshift (recap)

You met Amazon Redshift in Chapter 11 as AWS’s petabyte-scale data warehouse. Where Athena is ideal for occasional queries on files in S3, Redshift shines for frequent, complex, high-performance SQL analytics and business intelligence over large, structured data sets. Redshift Serverless removes cluster management, and Redshift Spectrum lets Redshift query data in S3 as well.

Amazon EMR

Amazon EMR is a managed platform for running open source big data frameworks such as Apache Spark, Apache Hadoop, Apache Hive, and Presto. It is the choice when a team needs to process very large data sets with these frameworks, for example machine learning feature engineering, log processing, or large-scale transformations, without building and tuning clusters by hand. EMR can run on EC2 clusters, on Amazon EKS, or as EMR Serverless.

AWS Lake Formation

AWS Lake Formation makes it easier to build, secure, and manage a data lake. Its most important job for the exam is centralized, fine-grained access control: you grant permissions on databases, tables, columns, and even rows in one place, and those permissions are enforced across Athena, Redshift Spectrum, EMR, and other integrated services. That way, the HR team can query salary columns while everyone else sees the same table without them.

Amazon QuickSight

Amazon QuickSight is a serverless, cloud-native business intelligence (BI) service for building interactive dashboards and reports. It connects to Athena, Redshift, RDS, S3, and many other sources, uses an in-memory engine called SPICE for fast performance, and scales to many users with pay-per-use pricing options. Generative AI features let users ask questions in plain language, such as “What were sales by region last quarter?”, and get a visual answer. AWS has been bringing QuickSight’s capabilities into a broader suite of business AI tools called Amazon Quick Suite, but the exam may still simply say “QuickSight.” Whenever a question says “visualize,” “dashboard,” or “BI,” think QuickSight.

Amazon SageMaker unified studio

As analytics and AI have converged, AWS introduced the next generation of Amazon SageMaker, centered on Amazon SageMaker Unified Studio: a single development environment where data engineers, analysts, and data scientists can find data, run SQL, build data pipelines, and develop ML and generative AI applications, using tools such as Athena, EMR, Glue, Redshift, and Bedrock from one place, with shared governance through a built-in catalog. Do not confuse it with Amazon SageMaker AI (Section 13.6), the service for building, training, and deploying ML models, which is now one of the tools available within the unified studio.

EMR), so each scales and is paid for independently. One copy of the data can serve many tools. Amazon SageMaker unified studio brings these tools together in one

workspace for data and AI teams.

Figure 13.1: A serverless data lake pipeline. Data is ingested into Amazon S3, cataloged with AWS Glue, queried with Athena, and visualized in QuickSight, with Lake Formation managing permissions.

Exam tip

Match the verb to the service. “Query data in S3 using SQL without loading it” is Athena. “Discover schema and catalog data” or “serverless ETL” is Glue. “Run Spark or Hadoop” is EMR. “Complex analytics in a data warehouse” is Redshift. “Dashboards” is QuickSight. “Fine-grained permissions on a data lake” is Lake Formation.

13.3 Streaming and search

Batch analytics answers yesterday’s questions. Some decisions cannot wait: flagging a fraudulent card payment, reacting to a spike in website errors, or updating a live leaderboard. That calls for streaming data, processed continuously as it arrives.

Amazon Kinesis Data Streams

Amazon Kinesis Data Streams captures and stores streaming data, such as clickstreams, application logs, and IoT telemetry, in real time. Data is written into shards, kept in order for a retention period you choose, and can be read by multiple consumer applications at once, for example an AWS Lambda function that raises fraud alerts and an Apache Flink application that computes live metrics. Because data is retained, consumers can replay it. You build or choose the consumers.

Amazon Data Firehose

Amazon Data Firehose (formerly Kinesis Data Firehose) is the easiest way to load streaming data into destinations. It is fully managed and scales automatically: you choose a destination such as Amazon S3, Amazon Redshift, Amazon OpenSearch Service, or supported HTTP endpoints and third-party tools, and Firehose buffers, optionally transforms or converts the data, and delivers it in near real time. There are no consumers to write. Firehose can read from a Kinesis data stream or receive data directly from producers.

Amazon Managed Service for Apache Flink and Amazon MSK

Amazon Managed Service for Apache Flink runs Apache Flink applications that analyze streaming data in real time, for example calculating rolling averages or detecting anomalies. Amazon Managed Streaming for Apache Kafka (Amazon MSK) is a fully managed service for Apache Kafka, a popular open source streaming platform. If a scenario says the company already uses Kafka and wants to stop managing Kafka clusters, the answer is MSK. Real-time streaming with Amazon Kinesis

Producers
Mobile app clicks
IoT sensors
Server logs
Kinesis Data Streams

shard 1: records...

shard 2: records...

ordered, replayable,
kept for a set period
Consumers (many per stream)
AWS Lambda: fraud alert in seconds
Managed Service for Apache Flink: live metrics
Amazon Data Firehose: delivery (below)
Firehose buffers, optionally transforms, and loads into:
Amazon S3 Redshift OpenSearch HTTP, partners

Data Streams: you build consumers; real-time; data can be replayed.

Data Firehose: fully managed delivery to a destination; near real time.

Amazon MSK: managed Apache Kafka, for teams already using Kafka.

Firehose can also receive data directly from producers, without Data Streams.

Figure 13.2: Producers write records to Kinesis Data Streams; multiple consumers process them in real time, and Data Firehose delivers data to storage and analytics destinations.

Common trap

Kinesis Data Streams and Data Firehose are easy to confuse. If the question emphasizes custom real-time processing by your own applications or replaying data, pick Data Streams. If it emphasizes simply delivering or loading streaming data into S3, Redshift, or OpenSearch with no administration, pick Data Firehose.

Amazon OpenSearch Service

Amazon OpenSearch Service makes it easy to deploy and run OpenSearch, an open source search and analytics suite (it also supports legacy Elasticsearch versions). Use it for full-text search (searching a product catalog with typo tolerance and relevance ranking), log analytics and observability (searching application logs in near real time), and dashboards through OpenSearch Dashboards. A serverless option removes cluster management, and its vector search capability is widely used in generative AI applications to find relevant documents.

13.4 AI, ML, and generative AI in plain English

These three terms are nested, like Russian dolls:

Artificial intelligence (AI) is the broad goal of building systems that perform tasks that normally need human intelligence, such as understanding language, recognizing images, or making decisions.

Machine learning (ML) is a way to achieve AI: instead of writing rules by hand, you train a model on example data so it learns patterns. Using the trained model to make predictions on new data is called inference. A spam filter trained on millions of labeled emails is classic ML.

Deep learning is a subset of ML that uses large neural networks; it powers modern image, speech, and language systems.

Generative AI is a subset of deep learning that creates new content: text, images, code, audio, and video. It is powered by foundation models (FMs), very large models pre-trained on vast amounts of data that can be adapted to many tasks. Large language models (LLMs) are foundation models specialized in language.

A few more terms appear in exam questions. A prompt is the input you give a generative model; prompt engineering is crafting prompts to get better results. Models read and write text in tokens, small chunks of words, and generative AI services usually charge per token. A hallucination is a confident-sounding answer that is false. Retrieval-augmented generation (RAG) reduces hallucinations by retrieving relevant passages from your own trusted documents and adding them to the prompt, so the model answers from your facts. Fine-tuning further trains a model on your own examples to adjust its behavior.

13.5 Generative AI on AWS: Amazon Bedrock and Amazon Q

Amazon Bedrock

Amazon Bedrock is a fully managed, serverless service for building generative AI applications with foundation models from leading AI companies and Amazon through a single API. You do not manage any infrastructure or host the models yourself. Models available include:

Anthropic Claude, a family of models known for reasoning, writing, coding, and analysis.

Amazon Nova, Amazon’s own family of models for text, images, video, and speech.

Meta Llama, openly available models.

Models from other providers, such as Mistral AI, Cohere, AI21 Labs, and Stability AI.

Because every model sits behind the same API, teams can try several, compare them, and switch as better models appear. Bedrock is built with privacy in mind: your prompts and data are not used to train the base models and are not shared with the model providers, and you can keep traffic private with VPC endpoints and encrypt data with your own keys. Bedrock’s main capabilities:

Playgrounds in the console for experimenting with prompts and model settings.

Knowledge Bases: a managed RAG capability. You connect data sources such as S3; Bedrock splits documents, creates embeddings, stores them in a vector store, and retrieves relevant passages at question time so answers are grounded in your content, with citations.

Agents: let a model plan and carry out multi-step tasks by calling your APIs and data sources, for example checking an order’s status and then booking a return. Amazon Bedrock AgentCore adds services for running and operating agents securely at scale.

Guardrails: configurable safeguards applied to inputs and outputs. You can block denied topics, filter harmful content, redact personal information (PII), and check that answers are grounded in source material to reduce hallucinations.

Model customization: fine-tuning and other techniques to adapt a model with your own data, which stays private to your account.

Model evaluation: compare models on your own tasks, automatically or with human reviewers.

Hands-on lab

Mini lab: try a foundation model in the Bedrock playground

1. In the AWS Management Console, open Amazon Bedrock in a Region where it is available.
2. Open Model catalog and read the card for a model such as Anthropic Claude or Amazon Nova. Some providers ask you to accept terms before first use.
3. In the navigation pane, choose Playgrounds, then Chat / Text, and select the model.
4. Enter a prompt, such as asking for a three-bullet summary of a short policy you paste in, and choose Run.
5. Lower the temperature setting and run again; notice the answer becomes more predictable. Raise it and notice more variety.
6. Look at the token counts and latency shown for the response; these drive cost and user experience.
7. Optionally, open Guardrails, create one that denies a topic (for example, investment advice), attach it in the playground, and test a prompt on that topic.

console (illustration) / Bedrock / Playgrounds / Chat
Console Services Search services Region: US East (N. Virginia)
Amazon Bedrock
Overview
Model catalog

Playgrounds

Chat / Text

Image

Build

Agents
Knowledge Bases
Flows

Safeguards

Guardrails

Tune and assess

customer email.

support assistant.

Guardrail
support-safety
Lower temperature =
more predictable
Compare mode lets you send the same prompt to two models side by

side to compare quality, latency, and cost before choosing one.

Figure 13.3: A simplified Bedrock chat playground with a model selector, prompt and response, and configuration settings such as temperature,

maximum output tokens, system prompt, and guardrail. (Illustration, not an actual AWS screenshot.) Amazon Q

Amazon Q is AWS’s family of generative AI assistants that are ready to use; you do not have to build anything. Amazon Q Business is an assistant for employees. It connects to company data sources, such as wikis, intranets, document stores, ticketing systems, and email, through built-in connectors, then answers questions, summarizes documents, and generates content based on that data. Crucially, it respects existing access permissions, so employees only get answers from content they are allowed to see. Administrators can also apply guardrails. Amazon Q Developer is an assistant for software developers and IT teams. In the IDE, command line, and AWS console it generates and explains code, writes tests, scans code for security vulnerabilities, helps upgrade applications (for example, to newer Java versions), and answers questions about your AWS resources and costs.

Exam tip

“Build our own generative AI application using foundation models via an API” is Bedrock. “Give employees a ready-made assistant that answers questions from our company documents” is Amazon Q Business. “Help developers write, debug, and secure code” is Amazon Q Developer.

13.6 Amazon SageMaker AI

Amazon SageMaker AI is a fully managed service to build, train, and deploy your own machine learning models at scale. It removes the heavy lifting from each step of the ML life cycle: preparing data, training on managed compute that is turned off when the job finishes, tuning, deploying models to scalable endpoints, and monitoring them in production. Choose it when pre-trained AI services cannot do the job and you need a custom model, for example predicting equipment failure from your own sensor history or scoring loan risk with your own data. Features you may see on the exam:

SageMaker Studio: a web-based environment (IDE) for the whole ML workflow.

SageMaker Canvas: a no-code, visual interface that lets business analysts build ML models and generate predictions without writing code.

SageMaker JumpStart: a hub of pre-trained models, including foundation models, and solution templates you can deploy and fine-tune quickly.

SageMaker Ground Truth: helps create high-quality labeled training data, using human labelers and automation.

SageMaker Clarify: detects potential bias in data and models and explains which features drove a prediction. SageMaker Model Monitor: watches deployed models for drops in data or prediction quality.

Bedrock or SageMaker AI? Bedrock gives you ready-made foundation models through an API with no infrastructure. SageMaker AI gives you full control to train and host any model, including your own, at the cost of more ML expertise.

The infrastructure underneath

All of this runs on AWS compute. For teams training or serving models themselves, AWS offers EC2 instances with powerful NVIDIA GPUs, and its own purpose-built chips: AWS Trainium for cost-efficient, high-performance model training, and AWS Inferentia for low-cost, high-throughput inference. You do not need details for the exam; just remember which chip goes with which phase.

The AWS AI and ML stack
more ready-made, less ML skill
more control and customization
Layer 3: AI applications and services (call an API, no model building)
Amazon Q Business
Amazon Q Developer
Rekognition Textract Comprehend
Transcribe Polly Translate Lex Kendra Personalize

Pre-trained models behind simple APIs, plus generative AI assistants.

Higher layers run on the lower ones; pick the highest layer that meets the need.

Figure 13.4: The AWS AI stack. Ready-made AI services sit on top, Bedrock and SageMaker AI in the middle, and ML infrastructure (GPU instances, Trainium, Inferentia) at the bottom.

13.7 Pre-trained AI services

At the top of the stack are AI services: pre-trained models behind simple APIs. You send an image, document, text, or audio file and get back a result, with no ML expertise required. The exam loves “which service?” questions here, and the trick is to start from the type of input.

Images and video

Amazon Rekognition analyzes images and video. It detects objects, scenes, and text in images; detects and compares faces (for example, for identity verification); flags inappropriate or unsafe content for moderation; recognizes celebrities; and tracks people and activities in video. Custom Labels lets you train it to recognize your own items, such as your company’s products or logos.

Documents

Amazon Textract extracts printed and handwritten text, forms (key-value pairs), and tables from scanned documents and images, going far beyond simple optical character recognition (OCR). Think invoices, loan applications, tax forms, and medical intake forms.

Amazon Kendra is an intelligent enterprise search service. It indexes documents across sources such as SharePoint, S3, and databases, and lets employees ask natural-language questions (“How many vacation days do new hires get?”) to get precise answers, not just a list of links. It can also serve as a retriever for generative AI applications.

Text and language

Amazon Comprehend uses natural language processing (NLP) to find meaning in text: sentiment (positive, negative, neutral, mixed), key phrases, entities such as people and places, language, topics, and personally identifiable information (PII). Companies use it to analyze product reviews, support tickets, and social media posts. Comprehend Medical extracts medical information from clinical text.

Amazon Translate provides fluent machine translation between many languages, for example to localize a website or translate customer chats in real time.

Speech and conversation

Amazon Transcribe converts speech to text. It is used for call center transcripts, meeting notes, subtitles, and searchable media archives, and can identify different speakers and redact sensitive information. A Transcribe Call Analytics option adds call-center insights.

Amazon Polly does the reverse: it turns text into lifelike speech in many voices and languages, for example to read articles aloud or give an app a voice.

Amazon Lex builds conversational interfaces (chatbots and voice bots) using the same technology family that powers Amazon Alexa. It understands the user’s intent, asks follow-up questions to collect details, and calls your business logic, for example to book an appointment. It integrates with Amazon Connect for contact-center bots.

Recommendations

Amazon Personalize creates real-time personalized recommendations from user behavior data such as clicks, views, and purchases, the kind of “customers who viewed this also viewed” experience Amazon.com is known for, without ML expertise.

Figure 13.5: Choosing a pre-trained AI service by input type: images and video, text, speech and voice, documents, and user behavior data.

Requirement Service
Moderate user-uploaded photos Rekognition
Pull fields from scanned invoices Textract
Gauge sentiment of reviews Comprehend
Subtitle recorded videos Transcribe
Read newsarticlesaloud Polly
Localizea website Translate
Build a booking chatbot Lex
Natural-language search of documents Kendra
“Recommended for you” on astore Personalize

Common trap

Transcribe and Polly are opposites: Transcribe is speech to text, Polly is text to speech. Similarly, Textract pulls text out of documents, while Comprehend understands the meaning of text you already have. A complete pipeline often uses both: Textract reads a scanned letter, then Comprehend detects its sentiment.

Real-world example

Real-world scenario: An insurance company receives thousands of claim calls and forms a day. Amazon Connect answers calls with a Lex bot that collects the policy number. Transcribe produces call transcripts and Comprehend flags angry customers for a supervisor. Textract extracts fields from uploaded claim forms, and Rekognition checks photos of vehicle damage. Adjusters use Amazon Q Business to ask questions about policy documents, while the data science team trains a custom fraud-scoring model in SageMaker AI. Each service does one job well, and none required the company to build its own speech or vision models.

13.8 Responsible AI basics

AI systems can make mistakes, reflect biases in their training data, or be misused. Responsible AI is the practice of designing, building, and operating AI systems so they are trustworthy and beneficial. AWS frames it around several dimensions:

Fairness: consider impacts on different groups of people and avoid unjustified bias.

Explainability: understand and evaluate why a system produced its output.

Privacy and security: protect data and models from misuse and unauthorized access.

Safety: prevent harmful output and misuse.

Controllability: have mechanisms to monitor and steer system behavior.

Veracity and robustness: produce correct outputs, even with unexpected or adversarial inputs.

Governance: apply best practices across the AI supply chain, including providers and deployers.

Transparency: help people understand when and how they are interacting with AI so they can make informed choices.

AWS provides tools that support these goals: Bedrock Guardrails to filter harmful content and reduce hallucinations, SageMaker Clarify to detect bias and explain predictions, SageMaker Model Monitor to catch drift, Amazon Augmented AI (Amazon A2I) to route low-confidence predictions to human reviewers, and AWS AI Service Cards, which document the intended uses, limitations, and design choices of AWS AI services.

Remember that responsible AI is also part of the shared responsibility model. AWS is responsible for the security of the underlying services and documents how they behave; the customer is responsible for choosing appropriate use cases, testing with their own data, keeping humans in the loop for high-stakes decisions, and complying with the laws that apply to them.

Good to know

AWS offers a separate certification, AWS Certified AI Practitioner, that goes deeper into generative AI and responsible AI. For CLF-C02, focus on recognizing services and the core ideas in this chapter.

Chapter 13 summary

Analytics on AWS follows ingest, store, catalog, query, and visualize; Amazon S3 is the usual data lake, separating storage from compute.

AWS Glue crawls and catalogs data and runs serverless ETL; Athena runs SQL directly on S3 and charges per data scanned; Redshift is the data warehouse; EMR runs Spark and Hadoop.

Lake Formation centralizes fine-grained data lake permissions; QuickSight builds BI dashboards; SageMaker Unified Studio brings data, analytics, and AI tools together.

Kinesis Data Streams captures real-time streams for custom consumers; Data Firehose loads streams into destinations; MSK runs managed Kafka; OpenSearch Service handles search and log analytics.

AI is the broad goal, ML learns from data, and generative AI creates new content using foundation models. Amazon Bedrock offers foundation models such as Anthropic Claude, Amazon Nova, and Meta Llama through one API, with Knowledge Bases (RAG), Agents, Guardrails, and customization.

Amazon Q Business is an employee assistant over company data; Amazon Q Developer helps developers code and operate on AWS.

SageMaker AI builds, trains, and deploys custom ML models; Trainium chips target training and Inferentia chips target inference.

Match AI services to input: Rekognition (images), Textract (documents), Comprehend (text meaning), Translate, Transcribe (speech to text), Polly (text to speech), Lex (chatbots), Kendra (search), Personalize (recommendations). Responsible AI covers fairness, explainability, privacy, safety, controllability, veracity, governance, and transparency, supported by tools like Guardrails, Clarify, and A2I.

Key terms
Term Meaning
Datalake Central store of raw datain any format
ETL Extract, transform, load
Glue Data Catalog Central metadatastore for datalake tables
Data warehouse Store optimized for complex analytics
Streaming data Data processed continuously as itarrives
Machine learning Training models to learn patterns from data
Inference Using atrained model to make predictions
Foundation model Large pre-trained model adaptable to many tasks
Prompt Input given to a generative model
Token Small unit of text that models processand bill on
Hallucination Plausible but incorrect model output
RAG Grounding answers in retrieved trusted content
Fine-tuning Further training a model on your own data
Guardrails Safeguards on generative AI inputsand outputs
Responsible AI Building AI that is fair, safe, and trustworthy
Practice questions

1. A company stores years of application logs as files in Amazon S3 and wants analysts to run occasional SQL queries on them without loading data or managing servers. Which service should it use?

A. Amazon Redshift provisioned cluster B. Amazon Athena
C. Amazon EMR on EC2 D. Amazon RDS

2. Which AWS service automatically discovers the schema of data in S3 and stores table definitions in a central catalog?

A. AWS Glue B. Amazon QuickSight
C. Amazon Kinesis Data Streams D. AWS Lake Formation

3. A data team runs Apache Spark and Hadoop jobs on premises and wants a managed AWS service for these frameworks. Which should it choose?

A. Amazon Athena B. Amazon EMR
C. AWS Glue DataBrew D. Amazon OpenSearch Service

4. Executives want interactive dashboards showing sales trends from data in Amazon Redshift. Which service should be used?

A. Amazon QuickSight B. Amazon Comprehend
C. AWS Glue D. Amazon Kendra

5. A company wants to grant analysts access to only certain columns and rows of tables in its S3 data lake, managed centrally across query services. Which service fits?

A. AWS Lake Formation B. Amazon Macie
C. AWS Glue crawlers D. Amazon Data Firehose

6. A company needs to capture website clickstream data in real time and have several custom applications process the same data, with the ability to replay it. Which service should it use?

A. Amazon Data Firehose B. Amazon Kinesis Data Streams
C. Amazon S3 Glacier D. Amazon QuickSight

7. A company wants the simplest fully managed way to load streaming IoT data into Amazon S3 and Amazon Redshift without writing consumer applications. Which service fits?

A. Amazon MSK B. Amazon Data Firehose
C. Amazon EMR D. AWS Glue Data Catalog

8. A company runs Apache Kafka clusters on EC2 and wants to reduce the operational burden while keeping Kafka APIs. What should it use?

A. Amazon Kinesis Data Streams B. Amazon MSK
C. Amazon SQS D. Amazon OpenSearch Service

9. A start-up wants to build a generative AI chatbot using foundation models such as Anthropic Claude and Amazon Nova through an API, without managing infrastructure. Which service should it use?

A. Amazon SageMaker Ground Truth B. Amazon Bedrock
C. Amazon Lex D. Amazon Comprehend

10. A company’s generative AI assistant sometimes gives made-up answers about internal policies. Which approach best grounds answers in the company’s own documents?

A. Increase the model’s temperature B. Use Amazon Bedrock Knowledge Bases (retrieval-augmented generation)
C. Use Amazon Polly D. Use AWS Inferentia instances

11. A bank’s Bedrock application must refuse to discuss investment advice and must redact customer personal information in responses. Which feature should it configure?

A. Bedrock Guardrails B. Bedrock Agents
C. SageMaker JumpStart D. Amazon Personalize

12. A company wants a ready-to-use generative AI assistant that answers employees’ questions using content from its wiki, document store, and ticketing system while respecting existing permissions. Which service fits?

A. Amazon Q Developer B. Amazon Q Business
C. Amazon Rekognition D. Amazon Textract

13. A data science team needs to build, train, and deploy a custom model to predict equipment failures from its own sensor history. Which service should it use?

A. Amazon SageMaker AI B. Amazon Translate
C. Amazon Transcribe D. Amazon Kendra

14. A social media app must automatically detect inappropriate content in images that users upload. Which service should it use?

A. Amazon Textract B. Amazon Rekognition
C. Amazon Comprehend D. Amazon Polly

15. A lender receives thousands of scanned loan application forms and needs to extract the form fields and tables automatically. Which service fits?

A. Amazon Textract B. Amazon Transcribe
C. Amazon Lex D. Amazon Personalize

16. Which pairings of requirement and service are correct? (Choose two.)

A. Convert call recordings into text: Amazon Transcribe B. Convert text into lifelike speech: Amazon Comprehend
C. Determine the sentiment of product reviews: Amazon

Comprehend

D. Build a chatbot: Amazon Polly
E. Translate text between languages: Amazon Kendra

17. A retailer wants to show “recommended for you” products based on each shopper’s browsing and purchase history, without ML expertise. Which service should it use?

A. Amazon Personalize B. Amazon Lex
C. Amazon Athena D. Amazon Rekognition

18. A company is concerned that its loan-approval ML model may treat applicants from different groups unfairly. Which AWS tool helps detect bias and explain model predictions?

A. Amazon SageMaker Clarify B. AWS Trainium
C. Amazon Data Firehose D. Amazon CloudWatch Logs

Answers and explanations

1. B. Athena is serverless and queries data in S3 with standard SQL without loading it. Redshift and EMR require more setup and management for occasional queries, and RDS is a transactional database.
2. A. AWS Glue crawlers infer schemas and store them in the Glue Data Catalog. QuickSight visualizes data, Kinesis streams data, and Lake Formation manages permissions on top of the catalog.
3. B. Amazon EMR is the managed platform for Spark, Hadoop, and other big data frameworks. Athena is SQL-only, DataBrew is visual data preparation, and OpenSearch is for search and log analytics.
4. A. QuickSight is AWS’s business intelligence service for dashboards. Comprehend analyzes text, Glue prepares data, and Kendra searches documents.
5. A. Lake Formation provides centralized, fine-grained permissions down to columns and rows across integrated services. Macie discovers sensitive data, crawlers catalog data, and Firehose delivers streams.
6. B. Kinesis Data Streams supports real-time ingestion, multiple custom consumers, and replay during the retention period. Firehose delivers data to destinations rather than serving multiple custom consumers with replay.
7. B. Data Firehose is fully managed delivery of streaming data into destinations such as S3 and Redshift. MSK requires Kafka consumers, EMR is for big data processing, and the Glue Data Catalog stores metadata.
8. B. Amazon MSK is managed Apache Kafka, so existing Kafka applications keep working. Kinesis Data Streams uses different APIs, and SQS is a message queue, not Kafka.
9. B. Bedrock provides foundation models from Anthropic, Amazon, Meta, and others through a single API with no infrastructure. Lex builds intent-based bots but is not the service for accessing foundation models; Ground Truth labels data.
10. B. Knowledge Bases implement retrieval-augmented generation, which retrieves trusted company content and adds it to the prompt. Raising temperature increases randomness, and Polly and Inferentia do not address grounding.
11. A. Guardrails can deny topics and redact personal information in inputs and outputs. Agents orchestrate tasks, JumpStart offers models, and Personalize makes recommendations.
12. B. Amazon Q Business connects to enterprise data sources and answers questions while respecting user permissions. Q Developer is for software development, and Rekognition and Textract analyze images and documents.
13. A. SageMaker AI is designed to build, train, and deploy custom ML models. The other services are pre-trained AI services for translation, transcription, and search.
14. B. Rekognition includes content moderation for images and video. Textract extracts document text, Comprehend analyzes text, and Polly produces speech.
15. A. Textract extracts text, form fields, and tables from scanned documents. Transcribe handles audio, Lex builds chatbots, and Personalize recommends items.
16. A, C. Transcribe converts speech to text, and Comprehend detects sentiment. Polly (not Comprehend) does text to speech, Lex (not Polly) builds chatbots, and Translate (not Kendra) translates.
17. A. Personalize generates personalized recommendations from user behavior data. Lex builds chatbots, Athena queries data, and Rekognition analyzes images.
18. A. SageMaker Clarify detects potential bias in data and models and explains predictions. Trainium is training hardware, Firehose delivers streams, and CloudWatch Logs stores logs.