Domain 4A: AWS AI/ML Services
Table of Contents
- AWS AI/ML Services
AWS AI/ML Services
Exam Domain: 4 — ML Implementation and Operations (20%) Task: Select appropriate AWS AI/ML managed services for a given use case
Why Managed AI Services Exist
Not every ML problem requires building a custom model from scratch. AWS AI Services represent the “buy” side of the build vs. buy decision — pre-trained models exposed as APIs.
ELI5: AWS AI Services are like ordering takeout vs. cooking from scratch. You don’t need a culinary degree to get a great meal — just call the right restaurant. Amazon Rekognition is the photo-recognition restaurant. Comprehend is the language restaurant. You call them, they return results. No kitchen required.
The Build vs. Buy Decision Tree
Do you have a well-defined problem with known solutions?
│
├─ YES → Does AWS offer a managed service for it?
│ ├─ YES → Does your domain require customization?
│ │ ├─ NO → Use the managed service directly (Rekognition, Comprehend, etc.)
│ │ └─ YES → Use the service's custom training feature (Custom Labels, Custom Entity)
│ └─ NO → Use SageMaker built-in algorithms or custom training
│
└─ NO → Research the problem space first
└─ Novel problem, proprietary data, strict latency/cost requirements?
└─ Build custom with SageMaker
Why this matters for the exam: Many questions give a scenario and ask which service to use. The discriminating factor is usually: Is it a standard task (face detection, sentiment)? Use managed AI. Is it domain-specific (medical defects, industry-specific entities)? Use custom training or SageMaker.
Vision Services
Amazon Rekognition
ELI5: Rekognition is like giving your application a pair of trained eyes. Show it a photo, it tells you what it sees — faces, objects, text, unsafe content — instantly, with no ML expertise required.
Image capabilities:
- Object and scene detection (labels with confidence scores)
- Facial analysis: age range, gender, emotions, face landmarks
- Face comparison: “Is this the same person?” across two photos
- Celebrity recognition: identifies public figures
- Text in image: OCR for naturally occurring text (signs, labels)
- PPE detection: hard hats, masks, gloves for workplace safety
- Content moderation: detect explicit/suggestive content
Video capabilities (adds on top of image):
- Person tracking: follow individuals across frames
- Activity detection: running, eating, waving
- Streaming video analysis (Kinesis Video Streams integration)
- Stored video analysis (asynchronous, S3-based)
Rekognition Custom Labels:
- Train your own image classifier on YOUR labeled images
- Transfer learning from Rekognition’s base model
- Fewer labels needed than training from scratch (few-shot learning)
- Use cases: industrial defect detection, proprietary product identification
Key discriminators:
| Scenario | Recommended Approach |
|---|---|
| Detect faces in photos | Rekognition (built-in) |
| Identify specific employees by face | Rekognition Face Collection + SearchFaces |
| Detect YOUR company’s specific products | Rekognition Custom Labels |
| Classify medical images (X-rays) | SageMaker Image Classification (custom) |
| Count objects in satellite imagery | SageMaker Object Detection (custom) |
Exam tip: If the question says “custom” or “proprietary” images with domain-specific categories, lean toward Custom Labels or SageMaker. If it says standard detection/faces/text, use base Rekognition.
NLP Services
Amazon Comprehend
Full-managed NLP. No training required for standard tasks.
Built-in capabilities:
- Sentiment analysis: Positive/Negative/Neutral/Mixed per document
- Entity recognition (NER): People, places, organizations, dates, quantities, events
- Key phrase extraction: Important phrases without full parsing
- Language detection: Identifies which language a document is in (100+ languages)
- Topic modeling: Latent Dirichlet Allocation (LDA) to cluster documents by topic
- PII detection: Identify and redact personally identifiable information
- Syntax analysis: Part-of-speech tagging
Custom Comprehend capabilities:
- Custom classification: Train your own document classifier (support tickets → priority level)
- Custom entity recognition: Train your own NER model (product codes, drug names, internal jargon)
Comprehend Medical:
- Healthcare-specific NER (clinical text, doctor notes, prescriptions)
- Detects: medications, dosages, conditions, anatomy, procedures, tests
- Relationship extraction: “Drug X” — “treats” — “Condition Y”
- ICD-10-CM and RxNorm ontology linking
- HIPAA-eligible
When to use what:
| Use Case | Service |
|---|---|
| Classify customer support tickets | Comprehend Custom Classification |
| Extract drug names from clinical notes | Comprehend Medical |
| Classify short product reviews by topic | Comprehend built-in + topic modeling |
| Train word embeddings for downstream tasks | SageMaker BlazingText (Word2Vec mode) |
| Custom chatbot intent classification | Lex OR SageMaker |
ELI5: Comprehend is like a linguist you can rent by the API call. It reads text and tells you what it means — the sentiment, the people mentioned, the topics discussed — without you needing to write a single line of NLP code.
Amazon Translate
Neural machine translation (NMT) service.
- Real-time translation: synchronous API, low latency
- Batch translation: translate large document sets via S3
- Custom terminology: ensure brand names, technical terms are translated consistently (e.g., “Amazon SageMaker” stays as-is)
- Active Custom Translation: fine-tune with your parallel corpus (your translations teach the model your style)
- Supports 75+ languages, auto-detect source language
- Use cases: multilingual customer support, content localization, real-time chat translation
Amazon Transcribe
Automatic Speech Recognition (ASR) — audio to text.
Core features:
- Streaming transcription (real-time) or batch (file in S3)
- Speaker diarization: “Who said what?” — labels different speakers
- Custom vocabulary: add domain-specific words (medical terms, product names)
- Vocabulary filters: redact/filter profanity or specific words
- Automatic punctuation and number formatting
- Multiple language support
Transcribe Call Analytics:
- Purpose-built for contact centers
- Speaker diarization (agent vs. customer)
- Sentiment analysis per speaker
- Call categorization: auto-tag calls by topic
- Issue detection, action item detection
- Turn-by-turn analysis
Transcribe Medical:
- Trained on medical speech, clinical terminology
- Primary care, cardiology, neurology, etc. specialty models
- HIPAA-eligible
- Use case: doctors dictating notes, clinical trial audio
Amazon Polly
Text-to-speech (TTS) — text to lifelike audio.
Voice types:
| Type | Quality | Cost | Use Case |
|---|---|---|---|
| Standard | Good | Lower | High-volume, cost-sensitive |
| Neural (NTTS) | Excellent, natural | Higher | Customer-facing, brand voice |
| Newscaster | Authoritative | Higher | News, podcasts |
| Conversational | Natural dialog | Higher | Chatbots, assistants |
Advanced features:
- SSML (Speech Synthesis Markup Language): control prosody, pace, emphasis, pauses
- Lexicons: pronunciation rules for domain-specific words (e.g., “AWS” → “A-W-S”)
- Speech marks: get timestamps for each word/sentence (for lip-sync or subtitle alignment)
- Bilingual voices (English + another language in one voice)
Amazon Lex
Conversational AI — build chatbots and voice interfaces.
Core concepts:
Bot
│
├─ Intent: What the user WANTS to do
│ Example: "OrderPizza", "CheckBalance", "BookFlight"
│
├─ Utterances: How users might EXPRESS an intent
│ Example: "I want a pizza", "Order me a large cheese pizza"
│
├─ Slots: Variables the bot needs to fill
│ Example: PizzaSize, PizzaType, DeliveryAddress
│
└─ Fulfillment: What happens when all slots filled
└─ Lambda function (calls backend, returns response)
ELI5: Lex is the brain behind Alexa. You define what your bot understands (intents), how users say it (utterances), what info it needs (slots), and what to do when it has enough info (Lambda fulfillment). Lex handles all the NLU complexity.
Lex V2 improvements over V1:
- Streaming conversations (multi-turn dialog)
- Better multi-language support in one bot
- Improved intent classification
- Visual conversation builder
When Lex vs. custom SageMaker:
- Use Lex: building chatbots, voice assistants, FAQ bots — standard NLU tasks
- Use SageMaker: highly specialized intent classification, multi-label intents, proprietary conversation data requires custom model
Integration patterns:
- Web/mobile: Amazon Lex SDK
- Contact center: Amazon Connect + Lex (auto-handles incoming calls)
- Slack, Facebook Messenger: built-in channel integrations
Amazon Textract
OCR++ — goes beyond pixel reading to understand document structure.
ELI5: Regular OCR is like a scanner that just takes a photo of text. Textract is like a smart assistant who reads your form, knows that “Name:” is a label and “John Smith” is the value, and returns that relationship to you — not just raw text.
Capabilities:
| Feature | API | What it returns |
|---|---|---|
| Raw text | DetectDocumentText | Lines, words, bounding boxes |
| Forms | AnalyzeDocument (FORMS) | Key-value pairs (label → value) |
| Tables | AnalyzeDocument (TABLES) | Table cells with row/col position |
| Queries | AnalyzeDocument (QUERIES) | Answers to specific questions |
| Signatures | AnalyzeDocument (SIGNATURES) | Signature detection |
| ID documents | AnalyzeID | Structured fields from IDs, passports |
| Expense/Invoice | AnalyzeExpense | Vendor, date, line items, totals |
How it differs from basic OCR:
- Understands document geometry and structure
- Maintains reading order even in multi-column layouts
- Extracts relationships, not just characters
- Works on scanned documents, PDFs, photos
Use cases: Invoice processing, form digitization, medical record extraction, mortgage document processing, compliance document review
Forecasting & Time Series
Amazon Forecast
AutoML-powered time series forecasting service.
ELI5: Forecast is like hiring a forecasting team that tries 6 different methods automatically and picks the best one for YOUR data. Instead of deciding whether to use ARIMA or neural networks, you upload data and Forecast figures it out. DeepAR in SageMaker is like hiring one specific expert who only uses deep learning.
How it works:
- Upload historical time series to S3
- Optionally add related time series (weather, price) and item metadata
- Forecast trains multiple algorithms, selects best via AutoML
- Generate forecasts for future periods with confidence intervals
Supported algorithms:
| Algorithm | Type | Best For |
|---|---|---|
| DeepAR+ | RNN/LSTM | Complex patterns, many series |
| Prophet | Additive model | Clear trends, holidays, seasonality |
| ARIMA | Statistical | Univariate, stationary series |
| ETS (Exponential Smoothing) | Statistical | Simple trends/seasonality |
| NPTS (Non-Parametric Time Series) | Non-parametric | Intermittent demand |
| CNN-QR | CNN + quantile regression | Complex patterns at scale |
Key data types:
- Target time series: what you’re forecasting (sales, demand, etc.)
- Related time series: inputs that influence the target (promotions, holidays, weather)
- Item metadata: static attributes about each item (category, size, brand)
Use cases: Demand forecasting, inventory planning, financial planning, workforce scheduling, energy/utility forecasting
When to use Forecast vs. SageMaker DeepAR:
- Forecast: need quick setup, want AutoML to try multiple algorithms, no ML expertise
- SageMaker DeepAR: need custom architecture, full control over training, specialized use case
Personalization & Recommendations
Amazon Personalize
Real-time personalization and recommendation engine.
ELI5: Personalize is Netflix’s recommendation engine as a managed service. Upload your user interaction data (“User 123 watched Movie A, User 123 bought Product B”), and Personalize learns each user’s preferences. Then ask “What should I show User 123?” and get a personalized ranked list back.
Recipes (algorithms):
| Recipe | What it does | Example use case |
|---|---|---|
| USER_PERSONALIZATION | Items ranked for a specific user | Homepage recommendations |
| POPULARITY_COUNT | Most popular items globally | Trending section |
| SIMILAR_ITEMS | Items similar to a given item | “More like this” |
| PERSONALIZED_RANKING | Re-rank a list for a user | Personalized search results |
| USER_SEGMENTATION | Segment users by behavior | Targeted marketing |
Key features:
- Event tracker: capture real-time user interactions (clicks, views, purchases) for immediate personalization updates
- Filters: rules to exclude items (already purchased, out of stock, age-restricted, wrong category)
- Cold-start handling: new users get popular items; new items get shown based on metadata similarity
- Contextual recommendations: incorporate context (device type, time of day, location)
- Batch recommendations: generate for all users offline
Integration: REST API, AWS SDK, or via Amazon Pinpoint for campaign targeting
When to use Personalize vs. SageMaker Factorization Machines:
- Personalize: e-commerce recommendations, media recommendations, managed service, quick time-to-value
- SageMaker FM: need custom loss functions, integrate into complex pipelines, academic/research context
Anomaly Detection Services
Amazon Lookout for Metrics
Detect anomalies in business metrics automatically — no ML expertise needed.
How it works:
- Connect to your metrics data source (Redshift, RDS, S3, CloudWatch, Salesforce, etc.)
- Lookout learns normal patterns (daily/weekly seasonality, trends)
- Alerts when metrics deviate from expected patterns
- Groups related anomalies into incidents
- Provides likely root causes
Use cases: Revenue drops, ad campaign performance anomalies, website traffic spikes, supply chain disruptions
Amazon Lookout for Equipment
Predictive maintenance from industrial sensor data.
- Analyzes vibration, temperature, pressure, flow rate from industrial equipment
- Learns normal operating patterns per machine
- Detects early warning signs of failure
- Reduces unplanned downtime
- Use case: manufacturing, energy, mining equipment monitoring
Amazon Lookout for Vision
Visual defect detection in manufacturing — computer vision for quality control.
- Upload labeled images of normal and defective products
- Trains a custom model on YOUR manufacturing images
- Real-time inspection at production speed
- Anomaly map: highlights WHERE the defect is
- On-premises inference via AWS Panorama integration
- Use case: PCB defect detection, product scratch detection, pharmaceutical pill inspection
Document & Knowledge Management
Amazon Kendra
Intelligent enterprise search — semantic search, not just keywords.
ELI5: Regular search engines (Ctrl+F) find exact words. Kendra understands meaning. Ask “What is the company vacation policy?” and it finds the right paragraph in your HR PDF even if it never says “vacation policy” verbatim — because it understands the semantic context.
Features:
- Natural language questions, not just keyword matching
- FAQ matching: exact question-answer pairs get highest ranking
- Document ranking: ML-based relevance scoring
- Incremental learning from user feedback
- Source attribution: tells you which document/paragraph answered the question
Data source connectors:
- S3 (documents, PDFs, Word, HTML)
- SharePoint, OneDrive
- Salesforce, ServiceNow
- Confluence, JIRA
- Relational databases
- Web crawlers (websites)
Use cases: Enterprise knowledge base, employee self-service portal, customer support deflection, compliance document search
Amazon Augmented AI (A2I)
Human review workflows for ML predictions.
ELI5: A2I adds a “human safety net” for your ML model. When the model is uncertain (low confidence), A2I routes that case to a human reviewer who makes the final call. Perfect for medical diagnoses, legal documents, or anywhere a wrong prediction is costly.
Built-in integrations:
- Amazon Rekognition (content moderation)
- Amazon Textract (forms extraction)
- Custom task types for any ML model
Human reviewer sources:
- Private workforce: your own employees (internal review, highest privacy)
- Vendor workforce: pre-screened third-party contractors
- Amazon Mechanical Turk: crowdsourced workers (for non-sensitive tasks)
Workflow:
- ML model makes prediction with confidence score
- If confidence < threshold → route to A2I human review queue
- Reviewer sees task in Labeling UI, makes decision
- Result returned; optionally used to retrain model
Use cases:
- Medical image review when AI confidence is low
- Financial document processing (compliance requires human sign-off)
- Content moderation with appeals process
Why this matters for the exam: A2I is the answer when a question says “human review required for low-confidence predictions” or “regulatory compliance requires human oversight.”
Developer & Other Specialized Services
Amazon CodeGuru
ML-powered code quality and performance.
CodeGuru Reviewer:
- Automated pull request code review
- Identifies: security vulnerabilities, resource leaks, input validation issues, concurrency bugs
- Languages: Java, Python
- Integrates with GitHub, Bitbucket, CodeCommit
CodeGuru Profiler:
- Application performance analysis (flame graphs)
- Identifies CPU bottlenecks, memory issues, expensive lines of code
- Recommendations to optimize performance
- Languages: Java, Python, Node.js, .NET, Go, Ruby
AWS Panorama
Computer vision at the edge — bring ML vision to existing IP cameras.
- Panorama Appliance: physical device connecting to existing cameras
- Run computer vision models locally without sending video to the cloud
- Low latency, offline capability, data privacy (video stays on-premise)
- Use cases: retail analytics (foot traffic), manufacturing quality inspection, facility security
- Deploy models from SageMaker or pre-built from AWS Marketplace
Amazon DevOps Guru
ML-powered operational insights for applications.
- Automatically detects operational anomalies in AWS accounts
- Correlates CloudWatch metrics, logs, events to find root cause
- Proactive: predicts resource exhaustion, performance degradation
- Recommendations: actionable fix suggestions
- Use case: SRE teams, reduce MTTR for production incidents
Amazon HealthLake
FHIR-compliant healthcare data lake with built-in NLP.
- Store, transform, and analyze health data at scale
- FHIR R4 compliant (Fast Healthcare Interoperability Resources)
- Medical NLP to extract insights from unstructured clinical text
- Integrated analytics with QuickSight, SageMaker
- Use case: clinical data warehousing, population health analytics
Amazon Monitron
End-to-end predictive maintenance for industrial equipment.
- Complete hardware + software solution (sensors + gateway + app)
- Monitors vibration and temperature via magnetic sensors
- ML-based anomaly detection with pre-trained models
- No ML expertise required — install and monitor
- Simpler than Lookout for Equipment (no custom data needed)
- Use case: pumps, motors, gearboxes, fans in factories
AWS DeepRacer, DeepLens (Educational)
| Service | Purpose | Who Uses It |
|---|---|---|
| AWS DeepRacer | Learn reinforcement learning via toy race car simulator | Developers learning RL |
| AWS DeepLens | Deep learning-enabled video camera | Developers learning computer vision |
Exam tip: These are educational tools, not production services. If a question asks about production ML at the edge, the answer is Panorama (for cameras) or IoT Greengrass + SageMaker Neo.
Master Service Selection Table
Why this matters for the exam: Scenario questions dominate Domain 4A. Memorize the “I want to…” → service mapping.
Vision & Image Tasks
| I Want To… | Use This Service |
|---|---|
| Detect faces and recognize emotions in photos | Rekognition |
| Compare two face photos to check if same person | Rekognition CompareFaces |
| Search a collection of faces (“find this person”) | Rekognition Face Collection |
| Detect PPE compliance in workplace photos | Rekognition PPE Detection |
| Moderate user-generated images for explicit content | Rekognition Content Moderation |
| Detect custom product defects in manufacturing photos | Rekognition Custom Labels or Lookout for Vision |
| Monitor factory cameras for defects in real time | Lookout for Vision + Panorama |
| Extract text from a scanned business card | Textract DetectDocumentText |
| Extract key-value pairs from a tax form | Textract AnalyzeDocument (FORMS) |
| Extract data from an invoice | Textract AnalyzeExpense |
Language & Text Tasks
| I Want To… | Use This Service |
|---|---|
| Detect sentiment in customer reviews | Comprehend |
| Find PII in documents and redact it | Comprehend PII detection |
| Classify support tickets into categories | Comprehend Custom Classification |
| Extract drug names from clinical notes | Comprehend Medical |
| Translate website content to 10 languages | Translate |
| Keep brand names unchanged in translations | Translate Custom Terminology |
| Convert customer call recordings to text | Transcribe |
| Analyze call center recordings for sentiment/issues | Transcribe Call Analytics |
| Transcribe doctor-patient conversations | Transcribe Medical |
| Convert article text to audio for podcast | Polly (Neural voice) |
| Build a customer service chatbot | Lex |
| Build a voice-activated ordering system | Lex + Connect |
| Search internal knowledge base with natural language | Kendra |
Forecasting & Recommendations
| I Want To… | Use This Service |
|---|---|
| Forecast product demand for inventory planning | Forecast |
| Recommend products to users based on past behavior | Personalize |
| Detect anomalies in business metrics (sales, clicks) | Lookout for Metrics |
| Detect equipment failure before it happens | Lookout for Equipment or Monitron |
| Build personalized email campaigns | Personalize + Pinpoint |
Oversight & Governance
| I Want To… | Use This Service |
|---|---|
| Add human review when model confidence is low | A2I (Augmented AI) |
| Detect bias in ML training data or model predictions | SageMaker Clarify |
| Get automated code reviews on pull requests | CodeGuru Reviewer |
| Find performance bottlenecks in application code | CodeGuru Profiler |
Comparison: Managed AI vs. SageMaker Custom
┌─────────────────────────────────────────────────────────────────┐
│ DECISION FRAMEWORK │
├─────────────────────┬───────────────────────────────────────────┤
│ MANAGED AI SERVICE │ SAGEMAKER CUSTOM │
│ (call the API) │ (train your own model) │
├─────────────────────┼───────────────────────────────────────────┤
│ Standard use case │ Domain-specific / proprietary task │
│ Out-of-box accuracy │ Accuracy on YOUR data more important │
│ No labeled data │ Have labeled training data │
│ Quick deployment │ Can invest in training pipeline │
│ Pay per API call │ Pay for training + inference instances │
│ Not customizable │ Full control over model architecture │
│ AWS manages model │ You manage model lifecycle │
└─────────────────────┴───────────────────────────────────────────┘
Exam tip: When a question mentions “pre-built” or “no training data available” or “needs to work out of the box” — managed AI service. When it says “custom” or “proprietary” or “domain-specific” — lean toward SageMaker or the custom tier of a managed service (Custom Labels, Custom Entity Recognition).