AWS Managed AI Services

Exam Domain: 2 — ML Model Development (26%) Task: Choose the right modeling approach — use managed services before building custom

When Managed vs Custom?

Decision Tree:
  "Can an AWS managed service solve this?"
       │
       YES → Use managed service (faster, cheaper, no ML expertise needed)
       │
       NO → "Can a SageMaker built-in algorithm solve this?"
              │
              YES → Use built-in algorithm
              │
              NO → Build custom model (bring your own container)

Exam tip: AWS always prefers the simplest, most managed solution. If a managed service can do it, that’s the answer.

Think of it this way: Choosing between managed services and custom models is like deciding whether to cook dinner. If you can buy a ready-made meal (managed service) that does the job, don’t spend hours cooking from scratch (custom model). AWS exam questions are designed to reward the person who reaches for the simplest tool first — the more managed, the better.

Natural Language Processing (NLP)

Amazon Comprehend

Purpose: Extract insights from text — no ML experience needed.

Comprehend Use Cases:
  Customer reviews   → Sentiment analysis
  Support tickets    → Topic modeling + classification
  Medical records    → Comprehend Medical (HIPAA-eligible)
  Legal documents    → Entity extraction + PII detection

ELI5: Comprehend is like having a super-fast reader processing thousands of reviews per second. Feed it a customer review and it reports back: “this person is angry (sentiment), they mention John Smith in New York (entities), they’re talking about returning a product (key phrases), and their phone number is in there (PII).” No ML expertise needed — it’s all pre-built.

Comprehend Medical — specialized for healthcare:

• Extract medical conditions, medications, dosages, procedures
• HIPAA-eligible
• Works with ICD-10-CM, RxNorm ontologies

Amazon Translate

• Neural machine translation (75+ languages)
• Real-time and batch translation
• Custom Terminology — force specific translations (brand names, technical terms)
• Active Custom Translation — fine-tune for domain-specific language
• Use case: website localization, multilingual chatbots

Amazon Transcribe

Purpose: Speech-to-text (ASR — Automatic Speech Recognition)

ELI5: Transcribe is Siri’s brain without Siri’s personality. It converts audio to text and can even figure out who said what (Speaker Diarization). Think of a call center: one recording, two speakers, and Transcribe produces a labeled transcript — “Agent: How can I help? Customer: My order is missing.” That labeled transcript can then feed into Comprehend for sentiment analysis.

Amazon Polly

Purpose: Text-to-speech (TTS)

• 60+ voices, 30+ languages
• SSML support — control pronunciation, pauses, emphasis, speed
• Lexicon — custom pronunciation for specific words
• Neural TTS (NTTS) — more natural, human-like speech
• Speech Marks — metadata for lip-syncing (viseme, word, sentence timings)
• Use case: accessibility, voice applications, IVR systems

Computer Vision

Amazon Rekognition

Purpose: Image and video analysis.

Rekognition Custom Labels:
  You provide:  10+ labeled images per class
  AWS handles:  Transfer learning, training, hosting
  You get:      Custom image classification API
  Use case:     "Is this a defective part?" (manufacturing QA)

ELI5: Rekognition is like a security camera that also has a film degree. It identifies faces, detects whether someone’s wearing a hard hat (PPE Detection), reads text on road signs, flags inappropriate content, and can even recognize celebrities. The “Custom Labels” feature lets you teach it your own categories with as few as 10 images — useful for things like spotting defective parts on a factory line.

Amazon Textract

Purpose: Extract text, forms, and tables from documents.

Textract vs Comprehend: Textract extracts text from images/PDFs. Comprehend analyzes the meaning of text.

ELI5: Regular OCR just “sees” characters — it reads “Name: John” as four separate characters. Textract understands document structure: it knows “Name” is a form field label and “John” is its value. So you get structured data back ({"Name": "John"}), not just a wall of text. That’s the difference between reading and comprehending a form.

Conversational AI

Amazon Lex

Purpose: Build chatbots and voice bots (same tech as Alexa).

Lex Concepts:
┌──────────────────────────────────────────┐
│  Bot                                     │
│  ├─ Intent: "OrderPizza"                 │
│  │   ├─ Utterances: "I want a pizza"     │
│  │   │              "Order me a pizza"    │
│  │   ├─ Slots: {size}, {topping}, {crust}│
│  │   └─ Fulfillment: Lambda function     │
│  │                                       │
│  └─ Intent: "CheckOrderStatus"           │
│      ├─ Utterances: "Where is my order?" │
│      ├─ Slots: {orderId}                 │
│      └─ Fulfillment: Lambda function     │
└──────────────────────────────────────────┘

• ASR (speech → text) + NLU (text → intent)
• Slot types: built-in (date, number) and custom
• Multi-turn conversation support
• Integrates with: Lambda, Connect (call center), Cognito (auth)

ELI5: Lex is literally Alexa’s brain — the same technology powers both. An “intent” is what the user wants to do (order a pizza). “Utterances” are different ways they might say it (“I want a pizza”, “Order me a pizza”, “Get me a large pepperoni”). “Slots” are the details you need to collect to fulfill it: size, topping, crust. Once all slots are filled, a Lambda function does the actual work.

Personalization & Recommendations

Amazon Personalize

Purpose: Real-time personalized recommendations.

Personalize Data Inputs:
  1. Interactions dataset (required): user-item interactions
  2. Users dataset (optional): user demographics
  3. Items dataset (optional): item metadata

Output: Ranked list of recommended items per user

Use cases: e-commerce product suggestions, content recommendations, personalized search

ELI5: Personalize is the “customers who bought this also bought…” engine. Netflix recommendations, Amazon product suggestions — same underlying collaborative filtering technology, now available as an API you call with a user ID and get back a ranked list of items. You bring your interaction history; AWS handles all the model training and hosting.

Amazon Forecast (retired — but may still appear)

• Time series forecasting using AutoML
• Built-in algorithms: DeepAR+, CNN-QR, Prophet, ARIMA, NPTS, ETS
• Use case: demand forecasting, inventory planning, financial forecasting

Note: Amazon Forecast was retired in 2024. For new projects, use SageMaker Canvas or DeepAR. May still appear in exam questions.

Document & Search Intelligence

Amazon Kendra

Purpose: Intelligent enterprise search using NLP.

• Semantic search — understands meaning, not just keywords
• Returns specific answers, not just document links
• Connectors: S3, SharePoint, Salesforce, databases, websites
• FAQ support — question-answer pairs
• Access control integration (who can see what)
• Use case: internal knowledge base search, customer self-service

Amazon Augmented AI (A2I)

Purpose: Human review workflow for ML predictions.

ML Prediction → Confidence Check
                    │
                    ├─ High confidence → Auto-approve
                    │
                    └─ Low confidence → Route to Human Reviewer
                                            │
                                            └─ Human reviews → Final decision

• Built-in integrations: Textract, Rekognition
• Custom: any ML model output
• Workforce: Mechanical Turk, private team, or third-party
• Use case: content moderation review, document verification

ELI5: A2I is the “phone a friend” lifeline for your ML model. When the model isn’t confident enough in its prediction, instead of guessing wrong, it routes the case to a human reviewer. Think of it like a junior doctor flagging uncertain X-rays for a specialist — automation handles the easy cases, humans handle the edge cases.

AI Assistants

Amazon Q Business

Purpose: Enterprise generative AI assistant.

Amazon Q Developer

Purpose: AI coding assistant.

AI Hardware

AWS Custom Chips

Instance mapping:
  Trainium   → ml.trn1 instances
  Inferentia → ml.inf1, ml.inf2 instances
  NVIDIA GPU → ml.p3 (V100), ml.p4d (A100), ml.g5 (A10G)

Quick Decision: Which Service?