AWS AI Services & Amazon Bedrock

Pre-built AI APIs — no ML expertise needed. Use when you don't want to build/train your own model.

---

AI Services Overview

graph TD
    AI[AWS AI Services] --> Vision[Computer Vision]
    AI --> NLP[Natural Language]
    AI --> Speech[Speech]
    AI --> Forecast[Forecasting]
    AI --> Rec[Recommendations]
    AI --> Search[Search]
    AI --> Codegen[Code Generation]
    AI --> GenAI[Generative AI]

    Vision --> Rek[Rekognition<br/>Images + Video]
    NLP --> Comp[Comprehend<br/>Text Analysis]
    NLP --> Tx[Textract<br/>Document OCR]
    NLP --> Trans[Translate]
    NLP --> Kendra[Kendra<br/>Intelligent Search]
    Speech --> Transcribe[Transcribe<br/>Speech → Text]
    Speech --> Polly[Polly<br/>Text → Speech]
    Forecast --> Fore[Amazon Forecast]
    Rec --> Pers[Personalize]
    Search --> Kendra
    GenAI --> Bed[Bedrock<br/>Foundation Models]
    Codegen --> CW[CodeWhisperer]

    style AI fill:#dbeafe,stroke:#3b82f6
    style GenAI fill:#f3e8ff,stroke:#9333ea

---

Amazon Rekognition

Computer vision API — images and video.

Key Use Cases:

• User verification (match selfie to ID)
• Content moderation (filter explicit images)
• Smart media archive (tag and search videos)
• Workplace safety monitoring (PPE compliance)

Exam tip: Rekognition Custom Labels — train Rekognition on your own images (few-shot) without deep ML knowledge.

---

Amazon Comprehend

NLP API — extract meaning from text.

Key Use Cases:

• Customer feedback analysis (sentiment)
• Legal document processing (entity extraction)
• Medical records (Comprehend Medical)
• GDPR compliance (PII detection + redaction)

Amazon Comprehend Medical

• Specialised NLP for healthcare and medical text
• Detects: diagnoses, medications, dosage, test results, anatomy
• Protected Health Information (PHI) detection for HIPAA compliance

---

Amazon Textract

Extract text and structured data from documents (beyond basic OCR).

Rekognition vs Textract:

• Rekognition = text in images (signs, photos)
• Textract = text in documents (PDFs, forms, tables)

---

Amazon Transcribe

Automatic Speech Recognition (ASR) — audio/video to text.

---

Amazon Polly

Text-to-Speech (TTS) — text to lifelike audio.

---

Amazon Translate

Neural machine translation — real-time or batch.

• 75+ languages
• Formality control — formal vs informal tone
• Custom terminology — preserve brand names, technical terms
• Active Custom Translation — fine-tune with parallel data (source → target examples)
• Real-time: API call with text
• Batch: S3 input → S3 output

---

Amazon Lex

Build conversational chatbots — powers Alexa.

• Supports voice + text
• Integrates with: Connect (call centres), Kendra (Q&A), Lambda (fulfilment)
• V2 supports multi-turn conversations and streaming responses

---

Amazon Kendra

Intelligent enterprise search — NLP-powered.

• Indexes documents from: S3, SharePoint, Salesforce, ServiceNow, RDS, websites
• Returns specific answers (not just documents) — understands questions
• Custom synonyms — map internal terminology
• Access Control — respects document-level permissions
• Incremental learning — improves from user feedback

Kendra vs OpenSearch:

• Kendra = semantic NLP-based question answering, enterprise docs
• OpenSearch = keyword/full-text search, log analysis, custom ranking

---

Amazon Forecast

Time series forecasting service — no ML expertise needed.

DeepAR (SageMaker) vs Amazon Forecast:

• Forecast = fully managed, no code, AutoML
• DeepAR = more control, custom training, SageMaker integration

---

Amazon Personalize

Real-time personalisation and recommendation service.

Concepts:

• Dataset Group — container for datasets
• Interactions dataset — required (user-item events: clicks, purchases, ratings)
• Items dataset — optional (item metadata: category, price)
• Users dataset — optional (user metadata: age, gender)
• Solution — trained model (choose recipe)
• Campaign — deployed solution for real-time recommendations

Real-time vs Batch:

• Real-time: Campaign endpoint — call API per user
• Batch: Batch Segment job — generate recs for all users at once → S3

---

Amazon Bedrock

Fully managed access to Foundation Models (FMs) via API.

Overview

• No infrastructure to manage — call FM via API
• Models from: Anthropic (Claude), Meta (Llama), Mistral, Amazon (Titan), Cohere, AI21
• Private — your data and prompts are NOT used to train base models
• Fine-tuned models stay in your account

Core Features

Knowledge Bases (RAG)

graph LR
    Docs[Your Documents<br/>S3] --> KB[Knowledge Base<br/>Bedrock]
    KB --> VS[Vector Store<br/>OpenSearch / Aurora / Pinecone]
    User[User Query] --> KB
    KB --> FM[Foundation Model]
    FM --> Answer[Grounded Answer]

• Manages chunking, embedding, and vector store indexing automatically
• Supported vector stores: OpenSearch Serverless, Aurora PostgreSQL, Pinecone, Redis Enterprise

Bedrock Agents

• Autonomously break tasks into steps, call APIs, retrieve knowledge
• Action Groups — define what APIs the agent can call (Lambda + OpenAPI spec)
• Memory — agents can remember conversation context across sessions
• Use for: customer service automation, data analysis, multi-step workflows

Guardrails

• Content filtering — block hate speech, violence, sexual content
• Denied topics — block specific subjects (e.g. competitor mentions)
• Word filters — block specific words/phrases
• PII redaction — detect and redact personal data
• Grounding check — verify response is grounded in context (reduces hallucination)

Amazon Titan Models

---

AI Services — Exam Comparison Table