Machine Learning
Managed ML platforms and pre-trained AI services for vision, speech, language, and forecasting.
Amazon Bedrock
Serverless API access to foundation models (Anthropic, Meta, Amazon, and others) for building generative AI apps without managing infrastructure.
- One API, many models — pick a foundation model per request; no servers, clusters, or ML expertise required
- Knowledge Bases = managed RAG (ground answers in your own data); Agents = multi-step task orchestration; Guardrails = content filtering
- Customize models with fine-tuning on your data — the base model provider never sees your data, and your data never trains the shared model
- Inference modes: on-demand (per-token), batch (async, ~50% cheaper), provisioned throughput (dedicated capacity, hourly)
Exam triggers
Phrases and scenarios that point here.
- "Generative AI", "foundation models", "build a chatbot/summarizer with the least operational effort" → Bedrock
- "Use company documents to answer questions" (RAG) → Bedrock Knowledge Bases
- "Prevent harmful or off-topic model responses" → Bedrock Guardrails
- Fully managed / serverless GenAI vs. building and training your own model → Bedrock, not SageMaker
Not this service
- Build, train, and deploy your own ML models with full control → SageMaker
- Ready-made AI assistant for business users or developers, no building required → Amazon Q
Use cases
- Internal support chatbot that answers from your company wiki — Knowledge Bases handles the retrieval so answers cite real policy, not model guesses
- Nightly batch job summarizing thousands of customer support tickets into themes — batch inference at half the on-demand price since nobody's waiting on it
- An agent that takes 'rebook my flight' and orchestrates the lookup, cancellation, and rebooking API calls as separate steps
- Fine-tuning a base model on your product catalog's tone and terminology so generated descriptions sound like your brand, not generic marketing copy
- Guardrails on a public-facing assistant so it refuses off-topic or unsafe prompts before they ever reach the model
Commonly confused with
Verified against AWS documentation · 2026-08-27
Amazon Comprehend
Managed NLP service that pulls meaning — sentiment, entities, language, PII — out of unstructured text.
- Send text, get back structured output: sentiment, entities, key phrases, dominant language, PII — no model to train
- Two modes: real-time APIs for one document or a batch of 25, and async jobs that read a whole corpus from S3
- Custom classification and custom entity recognition train on your own labels — you supply examples, not algorithms
- Limits: ~5,000 bytes per document on most real-time calls, 25 documents per batch, topic modeling is async-only
Exam triggers
Phrases and scenarios that point here.
- "Analyze customer feedback", "gauge sentiment", "extract insights from text", "no ML expertise on the team"
- Detect or redact PII inside free-form text before storing or sharing it → Comprehend, not Macie
- Scanned forms or PDFs → Textract first, Comprehend second — Comprehend only reads text, never images
- Call recordings or voice notes → Transcribe first, Comprehend second — audio is never Comprehend's input
Not this service
- Text trapped in scanned documents, forms, or images → Textract
- Sensitive data sitting at rest in S3 buckets → Macie
- Clinical notes, PHI, medication or ICD-10 extraction → Comprehend Medical
Use cases
- Support ticket triage — custom classification routes tickets to the right queue, sentiment flags the angry ones for escalation before the SLA clock runs down
- Scrubbing PII out of chat and email transcripts before they land in a data lake or get handed to an external analytics vendor
- Review mining at scale — entities plus key phrases across thousands of reviews to find which product features people actually talk about, and targeted sentiment to tell whether each one is praised or complained about
- Multilingual help desk — detect the incoming language, route to the right agent pool or hand off to Translate, all before a human sees the message
- Topic modeling over a legacy document archive nobody has read in years, to find out what's actually in there before setting a retention policy
Commonly confused with
Verified against AWS documentation · 2026-08-24
Amazon Lex
Managed service for building voice and text chatbots using the same speech recognition and language understanding that powers Alexa.
- A bot is intents (what the user wants), utterances (how they phrase it), and slots (the data needed to act on it)
- Same ASR + NLU engine as Alexa — one bot accepts speech or typed text, and a V2 bot holds multiple language locales
- Lambda hooks twice: dialog code hook to validate and steer mid-conversation, fulfillment hook to actually do the work
- Limits: 5-minute default idle session TTL (24h max); build on the draft version, publish immutable numbered versions, point aliases at them
Exam triggers
Phrases and scenarios that point here.
- "Chatbot", "conversational interface", "virtual agent", "understands natural language"
- "Same technology as Alexa" → Lex, every time
- Phone menu that understands what callers say instead of press-1 → Lex inside Amazon Connect
- Bot needs to look something up or write to a system → Lex intent fulfilled by Lambda
- One bot serving multiple languages → Lex V2 locales, not one bot per language
Not this service
- Speech to text only, no conversation → Transcribe
- Text to speech only → Polly
- Sentiment, entities, or key phrases out of text → Comprehend
- Natural-language search across documents → Kendra
- The contact centre itself (queues, agents, routing) → Amazon Connect; Lex is only the bot inside it
Use cases
- Order-status self-service — caller says "where's my delivery", Lex elicits the order number as a slot, Lambda queries DynamoDB, bot answers without an agent
- IVR front door in Amazon Connect that routes on intent instead of a nested keypad menu, and hands off to a human only when the fallback intent fires
- Internal helpdesk bot in Slack for password resets and laptop requests, with Lambda opening the ticket and returning the ticket ID
- Appointment booking where date, time, and service are slots — the dialog code hook rejects a closed-Sunday slot before the fulfillment hook ever writes to the calendar
- One support bot serving EN/ES/FR locales against a single Lambda backend, instead of three bots and three deployments to keep in sync
Commonly confused with
Verified against AWS documentation · 2026-08-24
Amazon Polly
Fully managed text-to-speech service that turns plain text or SSML into lifelike audio.
- Text in (plain or SSML), audio out — MP3, OGG Vorbis, PCM, or JSON speech marks
- Four engines behind one API: standard, neural, long-form, generative — quality and price rise in that order
- SynthesizeSpeech is real-time, 3,000 billed characters max; StartSpeechSynthesisTask is async, 100,000 characters, writes straight to S3
- Lexicons fix pronunciation, SSML controls pauses, pitch and rate — SSML tags are never billed
Exam triggers
Phrases and scenarios that point here.
- "Turn text into lifelike speech", "voice-enable the app", "read the article aloud" → Polly
- Audio → text is the opposite direction → Transcribe, not Polly
- Whole documents or text past the real-time limit → StartSpeechSynthesisTask + S3, with optional SNS on completion
- Brand names or acronyms pronounced wrong → lexicon; needs a pause or slower delivery → SSML
Not this service
- Speech → text (transcripts, call analytics) → Amazon Transcribe
- Working out what a caller wants and driving the conversation → Amazon Lex (Lex speaks using Polly)
Use cases
- IVR and contact centre prompts synthesized once and cached in S3 — changing a menu becomes a text edit and a redeploy, not a studio booking
- "Listen to this article" audio on a blog or news site, generated at publish time and served from CloudFront alongside the post
- E-learning narration on the long-form engine, so a 20-minute module doesn't drift into flat robot delivery halfway through
- Speech marks driving word-by-word highlighting or avatar lip-sync in reading and language-learning apps
- Voice replies for an LLM chatbot on the generative engine, streaming audio out while the model is still producing tokens
Commonly confused with
Verified against AWS documentation · 2026-08-24
Amazon Q
AWS's generative AI assistant family — Q Developer for coding and AWS tasks, Q Business for company knowledge.
- Two products, one name: Q Developer (code, IDE, console, CLI) and Q Business (answers from your company's docs and apps)
- Q Developer lives where you work — IDE, AWS Console, CLI — and is trained on AWS itself, so it can explain console errors and generate IaC
- Q Business connects to enterprise data sources (S3, SharePoint, Salesforce, 40+ connectors) and answers with citations, respecting user permissions
- Ready-made assistant, zero ML skills needed — you don't pick models, train anything, or manage infrastructure
Exam triggers
Phrases and scenarios that point here.
- "AI assistant for developers", "code suggestions", "explain this error in the console" → Q Developer
- "Employees ask questions about internal company documents" → Q Business
- "Generative AI" + "no ML expertise required" + "assistant" → Amazon Q, not SageMaker
- Upgrade/migrate legacy Java code automatically → Q Developer code transformation
Not this service
- Build your own generative AI app with model choice → Bedrock
- Enterprise search without the conversational assistant → Kendra
Use cases
- A junior engineer pastes a CloudFormation stack failure into the console chat and gets the actual cause instead of an hour on Stack Overflow
- Modernizing a decade-old Java 8 order-management app to Java 17 — the transformation agent does the mechanical upgrade work
- HR team asks 'what's the parental leave policy in Malaysia?' and Q Business answers from the policy PDFs in SharePoint, with the source linked
- Generating a first-draft Lambda function with correct SDK parameter formats, because Q is tuned specifically on AWS APIs
- Security scanning code as it's written, catching hardcoded credentials before they hit the repo
Commonly confused with
Verified against AWS documentation · 2026-08-27
Amazon Rekognition
Pre-trained computer vision API for finding objects, faces, text, and unsafe content in images and video.
- Pre-trained models behind an API — no training data, no model hosting; every result comes back with a confidence score you threshold yourself
- Image APIs are synchronous (call, get answer); Video APIs are async jobs — Start* → SNS notification → Get*, over video sitting in S3
- Face collections store face vectors, not photos — CompareFaces is 1:1, IndexFaces + SearchFacesByImage is 1:many
- Limits: JPEG/PNG only, 5 MB as raw bytes vs 15 MB from S3, stored video max 10 GB or 6 hours, 20 concurrent video jobs per account
- Streaming Video Analysis and Bulk Image Analysis closed to new accounts as of 2026-04-30 — existing users keep access, new builds use the Image APIs on extracted frames
Exam triggers
Phrases and scenarios that point here.
- "Detect objects/faces/text in images", "no ML expertise required", "pre-trained model"
- Filter explicit or suggestive user uploads → DetectModerationLabels
- Find every photo containing a specific person → face collection + face search, not repeated CompareFaces
- Text in a photo (sign, jersey, screenshot) → Rekognition; text in a scanned form or invoice → Textract
Not this service
- Documents, forms, tables, handwriting → Amazon Textract
- Objects the built-in label set doesn't know → Rekognition Custom Labels, or SageMaker if you need control over the model
- Speech inside the video → Amazon Transcribe
Use cases
- Screening profile photos at signup — S3 upload event triggers DetectModerationLabels, high confidence auto-rejects, borderline scores go to a human review queue instead of a guess
- Making a photo archive searchable by person — index each face once, then "show me every shot with this person" is a single search call rather than a full rescan
- Blurring faces and plates before CCTV footage lands in long-term storage, using the returned bounding boxes so the archived copy is already privacy-safe
- Reading text off things people photograph rather than scan — delivery labels, meter readings, membership cards, shelf tags
- Cataloguing a media library so editors search "stadium, crowd, night" instead of scrubbing through timelines
Commonly confused with
Verified against AWS documentation · 2026-08-24
Amazon SageMaker AI
Fully managed platform for building, training, and deploying your own ML models.
- Covers the whole ML lifecycle in one service: label, build in notebooks, train, tune, deploy, monitor
- You bring the data and the model — AWS brings managed compute. It is not a pre-trained API
- Four inference options: real-time (persistent endpoint), serverless (intermittent traffic), asynchronous (large payloads, scales to zero), batch transform (offline bulk)
- Limits: real-time 25 MB / 60s, serverless 4 MB / 60s, async 1 GB / 1 hour, batch transform runs GB-scale datasets for days
- Naming: the original service is now SageMaker AI. "Amazon SageMaker" now refers to the wider Unified Studio platform that wraps it
Exam triggers
Phrases and scenarios that point here.
- "Custom model", "train on our own data", "our own algorithm" → SageMaker, not Rekognition/Comprehend/Transcribe
- "No ML expertise", "business analyst", "no code" → Canvas; "automatically find the best model" → Autopilot
- Endpoint sits idle most of the day → Serverless Inference; large payload or minutes-long inference → Asynchronous Inference; no endpoint needed at all → Batch Transform
- "Reduce training cost" → Managed Spot Training (up to 90% off)
- "Accuracy degrades over time" → Model Monitor; "explain the prediction" or "detect bias" → Clarify
Not this service
- Off-the-shelf vision, speech, or text prediction → Rekognition, Transcribe, Comprehend
- Call a foundation model by API and pay per token → Bedrock
- Train and predict without leaving SQL → Redshift ML or Athena ML
Use cases
- Forecasting demand from five years of your own POS data, where no off-the-shelf model has ever seen your SKUs
- Hosting a fraud-scoring model behind a real-time endpoint that checkout calls inline on every transaction
- Async inference for a claims processor chewing through 200 MB scanned PDF bundles, scaling to zero overnight
- Batch transform to re-score the entire customer base for churn once a month — nothing left running between runs
- Managed Spot Training for overnight hyperparameter sweeps, where an interruption costs patience but not money
Commonly confused with
Verified against AWS documentation · 2026-08-26
Amazon Textract
Managed OCR that also extracts forms, tables, and handwriting as structured data.
- Goes past plain OCR — returns key-value pairs, table cells, checkboxes, and per-block confidence scores
- Sync (DetectDocumentText, AnalyzeDocument) for single pages; async (Start*/Get*) for multipage PDFs, S3-only, SNS on completion
- Pre-trained specialists: AnalyzeExpense for invoices/receipts, AnalyzeID for US licenses and passports — no templates to maintain
- Limits: 10 MB and 1 page sync; 500 MB and 3,000 pages async; async results expire after 7 days
Exam triggers
Phrases and scenarios that point here.
- "Extract text/data from scanned documents", "forms", "tables", "handwriting" → Textract
- Text in photos, signs, or product labels → Rekognition, not Textract
- Multipage PDF → async API + S3 + SNS; single image → sync API
- Low-confidence extractions need human sign-off → Textract + Amazon A2I
Not this service
- Text in photos and scenes rather than documents → Amazon Rekognition
- Meaning from the text — entities, sentiment, PHI → Amazon Comprehend
- Speech to text → Amazon Transcribe
Use cases
- Accounts payable: supplier invoices hit S3, AnalyzeExpense returns vendor, totals, and line items, and the approval workflow starts without anyone re-keying a number
- Insurance or benefits intake: scanned claim forms where the checkbox state matters as much as the text — selection elements tell you which boxes were actually ticked
- KYC and onboarding: AnalyzeID reads a licence photo to prefill the signup form, with the applicant confirming rather than typing
- Archive digitisation: decades of scanned PDFs run through an async job, output indexed into OpenSearch so a paper archive becomes searchable
- Lab reports and financial statements: table extraction keeps row/column relationships intact, so results load straight into a database schema
Commonly confused with
Verified against AWS documentation · 2026-08-24
Amazon Transcribe
Fully managed automatic speech recognition (ASR) that converts audio and video into text.
- Two modes: batch reads a media file from S3, streaming takes live audio over HTTP/2 or WebSocket
- Output is JSON — every word carries a timestamp and a confidence score; SRT/VTT subtitle files are optional extras
- Speaker partitioning (diarization) labels speakers sharing one channel; channel identification splits two-channel audio (agent vs caller)
- Accuracy tuning ladder: custom vocabulary (free word list) → custom language model (paid, trained on your own text)
- Limits: batch media under 4 hours and 2 GB, 100 concurrent jobs (up to 10,000 queued, FIFO), max 2 audio channels
Exam triggers
Phrases and scenarios that point here.
- "Make recordings searchable", "convert calls/meetings to text", "index audio content"
- Subtitles or closed captions for video → Transcribe, not Polly
- Live captioning or real-time agent assist → streaming, not batch
- Strip card numbers or SSNs out of call recordings → Transcribe PII redaction
Not this service
- Text → speech (the reverse direction) → Polly
- Sentiment, entities, or key phrases from the transcript → Comprehend
- Transcript in a different language → Translate; Transcribe only outputs the language that was spoken
Use cases
- Turning a backlog of recorded webinars and podcasts into searchable text, so someone can jump straight to the pricing discussion instead of scrubbing a timeline
- Contact centre QA — stereo call recordings split by channel so you can measure agent talk time and grep for required compliance phrases
- Auto-captioning pipeline for a video platform: upload lands in S3 → Lambda starts a job → SRT file written next to the asset
- Redacting PII from recordings before transcripts reach the data lake, so analysts can query them without ever touching card numbers
- Live captions for all-hands and training sessions, where remote or hard-of-hearing attendees need text as it's spoken
Commonly confused with
Verified against AWS documentation · 2026-08-24
Amazon Translate
Fully managed neural machine translation for text, across 75 languages.
- Text in, text out — source language is either declared or set to `auto` (which calls Comprehend behind the scenes)
- Two modes: real-time (TranslateText, TranslateDocument) and asynchronous batch over an S3 folder
- Customization without training a model: custom terminology pins individual terms, parallel data (ACT) steers domain and style
- Limits: 10,000 bytes per real-time text call, 100 KB per real-time document, 20 MB per batch document
Exam triggers
Phrases and scenarios that point here.
- "Translate", "localize", "multilingual", "in the customer's preferred language"
- Thousands of documents already sitting in S3 → batch translation job, not a loop of real-time calls
- Brand or product names must survive untranslated → custom terminology
- Audio anywhere in the requirement → Transcribe before, Polly after — Translate only handles text
Not this service
- Speech to text → Amazon Transcribe. Text to speech → Amazon Polly
- Detect language, sentiment, or entities without translating → Amazon Comprehend
Use cases
- Localizing a support knowledge base: nightly batch job reads an S3 prefix of HTML articles, writes translated copies to an output prefix in ten target languages — the job runs under an IAM data access role, not the caller's credentials
- Live agent chat where the rep types English and the customer reads Thai — a single TranslateText call sits inline in the message path, fast enough that neither side notices
- Normalizing multilingual product reviews to English before analytics, so the sentiment pipeline downstream only ever has to reason about one language
- Subtitle pipelines: Transcribe produces the transcript, Translate produces the caption tracks, Polly optionally voices them — three services, one chain
- Catalogue localization where custom terminology locks model names, SKUs, and legal disclaimers so "Nimbus Pro" never ships as "Cloud Professional"
Commonly confused with
Verified against AWS documentation · 2026-08-24