AWS Machine Learning Services - AI Without Building Models

What you will learn

• The two categories of AWS ML — pre-built services vs custom model training
• Amazon Rekognition — image and video analysis without any ML knowledge
• Amazon Transcribe — speech to text with PII redaction
• Amazon Polly — text to speech in multiple languages and voices
• Amazon Translate — real-time text translation
• Amazon Comprehend — natural language understanding and sentiment analysis
• Amazon Lex — building conversational chatbots (the engine behind Alexa)
• Amazon Kendra — intelligent document search powered by ML
• Amazon Textract — extracting text and data from scanned documents
• Amazon Personalize — real-time personalised recommendations
• Amazon SageMaker — when you need to train your own custom models

Why this matters

A Hotstar content moderation team used to manually review flagged videos — slow, expensive, and inconsistent. Rekognition now scans every uploaded video automatically, flags inappropriate content in seconds, and routes only genuinely borderline cases to human reviewers. At Razorpay, customer support used to search through PDFs manually for transaction details. Textract now extracts data from uploaded bank statements and Kendra makes that data searchable instantly. At Swiggy, the recommendation engine that suggests restaurants and dishes is powered by Amazon Personalize — trained on order history, updated in real time, without a single ML engineer maintaining it. These are not experimental projects. They are production systems built on AWS ML services that any developer can integrate in a day.

Two Categories of AWS ML

AWS ML services split into two clean categories:

Start with pre-built services. Only reach for SageMaker when the pre-built services cannot solve your problem.

Amazon Rekognition — See What Is in Images and Videos

Rekognition analyses images and videos and tells you what it sees — objects, faces, text, activities, and whether content is appropriate.

What it can detect:

Real use case at Hotstar:

Amazon Transcribe — Speech to Text

Transcribe converts audio to text. Give it an audio or video file and it returns a text transcript with timestamps, speaker identification, and punctuation.

Key features:

PII Redaction — important for compliance:

Transcribe can automatically detect and redact personally identifiable information from transcripts.

Critical for call centre recordings, customer support audio, and any audio containing personal data.

Amazon Polly — Text to Speech

Polly converts text into lifelike speech. Give it text. Get back an audio file in dozens of voices, languages, and accents.

Voices:

SSML — Speech Synthesis Markup Language:

Control exactly how Polly speaks — pauses, emphasis, speaking rate, pitch.

Use cases:

• Voice responses in IVR systems
• Accessibility features (read page content aloud)
• Audio content from text articles
• Language learning applications

Amazon Translate — Real-Time Translation

Translate converts text from one language to another. Supports 75+ languages.

Auto-detect source language: Translate identifies the source language automatically — you do not need to tell it what the input language is.

Use cases:

• Translate user-generated content in real time (reviews, comments, messages)
• Multilingual customer support (translate incoming messages, outgoing responses)
• Localise application content for different regions

Amazon Comprehend — Understand Text

Comprehend reads text and extracts meaning from it — sentiment, key phrases, entities, language, and topics.

What Comprehend can do:

Amazon Comprehend Medical:

A specialised version that understands medical text — clinical notes, prescriptions, discharge summaries. Extracts: medications, dosages, diagnoses, procedures, and Protected Health Information (PHI).

Amazon Lex — Conversational Chatbots

Lex is the same technology that powers Amazon Alexa. It understands natural language and maintains conversation context to handle multi-turn dialogues.

Key concepts:

Lex + Connect:

Amazon Connect is AWS's contact centre service. Lex powers the automated voice responses — the IVR bot that handles simple requests automatically before routing complex ones to human agents.

Amazon Kendra — Intelligent Document Search

Kendra indexes your documents — PDFs, Word files, FAQs, web pages, databases — and makes them searchable with natural language questions.

Connectors for: S3, SharePoint, Salesforce, Confluence, ServiceNow, RDS, OneDrive. Index once. Search everything.

Use case at Razorpay: legal and compliance team searches thousands of contracts, regulations, and internal policies. Kendra finds the exact clause they need in seconds.

Amazon Textract — Extract Data from Documents

Textract goes beyond simple OCR. It does not just read text — it understands the structure of documents and extracts data in a meaningful way.

What Textract can extract:

Use case: Razorpay KYC process. Users upload scanned Aadhaar card. Textract extracts name, date of birth, address, and Aadhaar number automatically. No manual data entry.

Amazon Personalize — Real-Time Recommendations

Personalize creates recommendation models from your user interaction data and returns real-time personalised recommendations via API call.

No ML knowledge required. You provide the data. Personalize handles the model training, hosting, and serving.

Use cases: restaurant recommendations (Swiggy), product recommendations (e-commerce), content recommendations (Hotstar), "customers also bought" (retail).

Amazon SageMaker — Train Your Own Models

When pre-built services cannot solve your problem — your data is unique, your domain is specialised, accuracy requirements are very high — SageMaker lets you build and train custom ML models.

SageMaker is significantly more complex and expensive than the pre-built services. It requires ML expertise. Use it only when pre-built services fall short.

Hands-on Lab — Rekognition and Comprehend in the Console

Step 1 — Test Rekognition image analysis

Step 2 — Test content moderation

Step 3 — Test Comprehend sentiment analysis

Step 4 — Test entity recognition

Step 5 — Test Translate

Step 6 — Test Textract

No cleanup needed — these are all console demos with no persistent resources created.

Common Mistakes to Avoid

Common Mistake

Jumping to SageMaker before trying pre-built services. SageMaker requires ML expertise, data preparation, model training time, and ongoing maintenance. Rekognition, Comprehend, Textract, and Personalize solve most common AI use cases via a simple API call. Try the pre-built service first. Use SageMaker only when the pre-built service genuinely cannot solve your problem.

Common Mistake

Using Rekognition facial recognition without understanding the compliance requirements. In some countries and jurisdictions, storing and comparing facial data requires explicit user consent and has strict regulatory requirements. Check your local laws before building systems that identify individuals by face.

Tip

All the pre-built AI services (Rekognition, Transcribe, Polly, Translate, Comprehend, Lex, Kendra, Textract, Personalize) can be combined. A customer support call → Transcribe converts audio to text → Comprehend extracts sentiment and key phrases → Translate converts to English for a global team → stored and searched by Kendra. Each service does one job and they chain together naturally.