Task 2.1: ML Model and Foundation Model (FM) Development - Choose appropriate modeling approaches for ML and AI solutions

Skill 2.1.8: Apply AWS AI services to solve specific business problems (for example, Amazon Textract, Amazon Rekognition, Amazon Comprehend, Amazon Transcribe)

Overview

This skill tests your ability to identify when a managed AWS AI service is the correct modeling approach instead of a custom ML model, a foundation model, or simple rule-based logic. The core principle is:

When a task is already solved as a standard capability, call the managed service first.

AWS AI services are pre-trained, fully managed, and production-ready. They remove the need to gather labeled data, train a model, tune hyperparameters, and manage hosting infrastructure. For many business problems involving documents, images, video, text, and speech, a managed service is the fastest, cheapest, and most reliable option.

The four services most commonly tested in the MLA-C02 exam are:

AWS Service	Key business problem it solves	Primary input
Amazon Textract	Extract text, forms, tables, invoices, and identity fields from scanned documents	PDF, JPEG, PNG, TIFF
Amazon Rekognition	Analyze images and video for objects, faces, moderation labels, and text	Images, stored video, streaming video
Amazon Comprehend	Understand structure, sentiment, entities, key phrases, and PII in text	Raw text documents
Amazon Transcribe	Convert speech from audio recordings into text	Audio files, streaming audio

Where Managed AI Services Fit in the Modeling Approach Decision

Before selecting a modeling approach, evaluate where the answer comes from.

flowchart TD
    A[Business Problem] --> B{Is the logic fixed and rule-based?}
    B -- Yes --> C[Use plain code or simple rules]
    B -- No --> D{Is this a standard, pre-defined AI task?}
    D -- Yes --> E[Use a managed AWS AI service]
    D -- No --> F{Does the pattern exist only in your proprietary data?}
    F -- Yes --> G[Train a custom ML model with SageMaker AI]
    F -- No --> H[Use a foundation model in Amazon Bedrock]
    E --> E1[Textract: Document OCR]
    E --> E2[Rekognition: Image/Video analysis]
    E --> E3[Comprehend: Text NLP]
    E --> E4[Transcribe: Speech-to-text]

Key takeaway: If the task is standard, such as extracting text from a PDF, identifying objects in an image, detecting sentiment in a review, or transcribing a call, do not build a custom model. Call the managed AI service.

Deep Dive: Amazon Textract

Amazon Textract extracts text, handwriting, forms, tables, invoices, and identity fields from scanned documents.

Core APIs and Their Business Use Cases

API	What it does	Business problem
`DetectDocumentText`	Basic OCR: lines, words, bounding boxes	Simple text extraction from digital documents
`AnalyzeDocument`	Forms: key-value pairs and checkboxes	Loan applications, patient intake forms, tax forms
`AnalyzeDocument`	Tables: table structure with rows/columns	Financial statements, compliance reports
`AnalyzeExpense`	Invoice and receipt fields: vendor, tax, total, line items	Accounts payable automation, expense management
`AnalyzeID`	Identity fields: name, DOB, ID number, address	Customer onboarding, KYC, age verification

Input and Output Characteristics

Input formats: PDF, JPEG, PNG, TIFF
Data sources: Amazon S3 or direct byte input
Output: JSON with text, confidence scores, page numbers, bounding boxes, relationships between form keys and values, and table cells
Processing: Supports synchronous single-document and asynchronous batch processing

Typical Scenario

A shipping company receives thousands of scanned delivery invoices daily. The finance team manually enters vendor name, invoice number, total amount, and line items into SAP.

Best approach: Use Amazon Textract AnalyzeExpense to automatically extract invoice fields and line items. This requires no training data and is far faster than building a custom computer vision model.

Deep Dive: Amazon Rekognition

Amazon Rekognition analyzes images and video for objects, scenes, faces, text, and unsafe content.

Core Capabilities

Capability	What it does	Example business problem
Label detection	Identifies objects, scenes, activities, and concepts	Auto-tagging product images in e-commerce
Face detection	Detects faces and attributes: age range, emotions, landmarks	Attendee sentiment in physical stores
Face search	Compares a face against a stored face collection	Building access control, identity verification
Content moderation	Detects explicit, violent, or suggestive content	Social media profile photo moderation
Text detection	Detects text in natural scenes, signs, and labels	Reading license plates, product labels
Custom Labels	Trains custom object or scene detection from your own labeled data	Detecting proprietary product defects
Video analysis	Asynchronous analysis of stored video for labels, faces, people, and paths	Security camera footage review

Image vs. Video

Image analysis: Synchronous API, low latency, ideal for real-time moderation or search.
Video analysis: Asynchronous API for stored video in S3, or streaming with Kinesis Video Streams for near-real-time detection.

Important Distinction

Amazon Rekognition can detect text in images, but Amazon Textract is the correct service for extracting large amounts of text from documents. Rekognition text detection is intended for short text in natural scenes, signs, or product packaging.

Deep Dive: Amazon Comprehend

Amazon Comprehend performs natural language processing (NLP) on raw text to find structure, sentiment, entities, key phrases, and personally identifiable information (PII).

Core APIs

API	What it does	Example business problem
`DetectDominantLanguage`	Identifies the language of a document	Routing multilingual support tickets
`DetectEntities`	Finds people, places, organizations, dates, quantities	Extracting company names from news articles
`DetectKeyPhrases`	Extracts important phrases from text	Summarizing customer reviews
`DetectSentiment`	Returns positive, negative, neutral, or mixed sentiment	Monitoring social media brand perception
`DetectPiiEntities`	Identifies PII: name, SSN, email, bank account, address	Compliance with data privacy policies
`DetectSyntax`	Analyzes parts of speech and sentence structure	Preparing text for downstream processing
Topic modeling	Discovers themes across large document sets	Grouping support tickets by root cause
Custom classification	Classifies documents into your own categories	Sorting emails into department-specific queues
Custom entity recognition	Detects custom terms such as part numbers or product codes	Extracting proprietary product IDs from service logs

Batch vs. Real-Time

Real-time: Single document synchronous API for quick NLP tasks.
Batch: Asynchronous jobs for large document sets, topic modeling, and PII scanning at scale.

Typical Use Case

A customer support team wants to automatically route inbound messages to the correct department and tag the emotional tone of each message.

Best approach: Use Amazon Comprehend DetectSentiment to gauge tone and DetectKeyPhrases or custom classification to route by topic.

Deep Dive: Amazon Transcribe

Amazon Transcribe converts speech from audio recordings into written text with timestamps, speaker labels, and confidence scores.

Core Modes and Features

Feature	What it does	Business problem
Batch transcription	Transcribes audio files stored in S3 asynchronously	Transcribing meeting recordings, podcasts
Streaming transcription	Transcribes real-time audio over WebSocket or HTTP/2	Live call captioning, virtual assistants
Automatic language detection	Identifies the spoken language automatically	Multilingual customer support
Custom vocabularies	Improves accuracy for domain-specific terms	Transcribing medical terminology
Custom vocabulary filters	Removes or masks unwanted words	Meeting notes without profanity
Speaker diarization	Distinguishes between multiple speakers	Interview and meeting transcripts
Channel identification	Separates audio from different channels	Call center agent vs. customer audio
Call Analytics	Provides sentiment, categories, redaction, and issue detection	Contact center quality monitoring
Subtitle generation	Outputs WebVTT or SRT subtitle files	Video captioning and accessibility

Input and Output

Input formats: MP3, MP4, WAV, FLAC, OGG, WebM
Output: JSON transcript with word-level timestamps, confidence scores, speaker labels, and punctuation.

Integration with Amazon Comprehend

A common pattern is:

flowchart LR
    A[Audio file in S3] --> B[Amazon Transcribe]
    B --> C[Transcript JSON in S3]
    C --> D[Amazon Comprehend]
    D --> E[Sentiment + entities + key phrases]
    E --> F[Business dashboard or alerting]

This pipeline solves tasks such as:

Contact center sentiment analysis
Sales call key topic extraction
Compliance monitoring for risky language

Service Selection Decision Guide

When the exam asks which AWS AI service to use, map the data type and task to the correct service.

Business problem	Input	Required output	Best service
Extract total amount and vendor from invoices	Scanned PDF invoices	Structured JSON fields	Amazon Textract AnalyzeExpense
Extract form key-value pairs from loan applications	Scanned forms	JSON key-value pairs	Amazon Textract AnalyzeDocument
Detect unsafe images before displaying them	User-uploaded images	Moderation labels	Amazon Rekognition
Identify celebrities in a video archive	Stored video	Celebrity names and timestamps	Amazon Rekognition video
Check if a given face matches an employee badge photo	Two images	Similarity score	Amazon Rekognition CompareFaces
Determine customer mood from support chat	Raw chat text	Sentiment score	Amazon Comprehend
Find all bank account numbers in documents	Text document	PII entity locations and types	Amazon Comprehend
Convert voicemail audio into readable text	Audio file	Text transcript	Amazon Transcribe
Label speaker and sentiment in a call recording	Audio file	Transcript + speaker labels + sentiment	Transcribe Call Analytics or Transcribe + Comprehend
Extract text from an image of a street sign	Image	Short text string	Amazon Rekognition DetectText
Extract hundreds of pages from a scanned report	Multi-page PDF	Full text + tables	Amazon Textract

Common Multi-Service Solutions

Many real-world exam scenarios combine multiple AI services.

Contact Center Intelligence Pipeline

flowchart TD
    A[Call audio from S3] --> B[Amazon Transcribe]
    B --> C[Speaker-separated transcript]
    C --> D[Amazon Comprehend Sentiment]
    C --> E[Amazon Comprehend PII detection]
    D --> F[Agent scorecard]
    E --> G[Redacted transcript for CRM]

Result: The call center automatically scores agent tone, detects sensitive customer information, and stores a compliant transcript.

Document Processing and Understanding

flowchart TD
    A[Invoices in S3] --> B[Amazon Textract]
    B --> C[Structured invoice JSON]
    C --> D[Amazon Comprehend Entities]
    D --> E[Vendor names and locations]
    C --> F[Validation rules]
    F --> G[Accounts payable automation]

Scenario-Based Exam Examples

Scenario 1

A financial services company must process scanned customer identification documents such as passports and driver's licenses. The solution must extract name, date of birth, document number, and expiration date without training a custom model.

Which AWS service should the company use?

A. Amazon Rekognition DetectFaces
B. Amazon Textract AnalyzeID
C. Amazon Comprehend DetectEntities
D. Amazon Transcribe

Correct Answer: B. Amazon Textract AnalyzeID

Explanation:
Amazon Textract AnalyzeID is purpose-built to extract key fields from government-issued identity documents. It requires no custom training and produces structured JSON. Amazon Rekognition is for visual face analysis, not document data extraction. Amazon Comprehend works with raw text, not image-based documents. Amazon Transcribe converts speech to text.

Scenario 2

A media platform needs to automatically detect and block images containing nudity, violence, or drug-related content before the images are shown to users. The solution must be low latency and require no manual label review.

Which AWS service meets these requirements?

A. Amazon Rekognition moderation labels
B. Amazon Comprehend sentiment analysis
C. Amazon Textract object detection
D. Amazon Transcribe content filtering

Correct Answer: A. Amazon Rekognition moderation labels

Explanation:
Amazon Rekognition provides pre-trained moderation APIs that detect explicit, suggestive, violent, and drug-related visual content. It operates synchronously on images, making it suitable for real-time moderation. Amazon Comprehend and Amazon Transcribe cannot analyze images. Amazon Textract does not perform moderation.

Scenario 3

A contact center stores customer call recordings in Amazon S3. The company wants to extract the main discussion topics, identify whether customers were frustrated, and generate a searchable transcript.

Which combination of AWS services is MOST appropriate?

A. Amazon Rekognition + Amazon Textract
B. Amazon Transcribe + Amazon Comprehend
C. Amazon Textract + Amazon Comprehend
D. Amazon Transcribe + Amazon Rekognition

Correct Answer: B. Amazon Transcribe + Amazon Comprehend

Explanation:
Amazon Transcribe converts the call audio into text. Amazon Comprehend then analyzes that transcript to find entities, key phrases, and sentiment. This is the standard architecture for contact center intelligence. Rekognition and Textract are used for images and documents, not audio.

Exam Tips and Traps

Tips

Match the service to the data type first.
Documents → Textract
Images or video → Rekognition
Text → Comprehend
Audio → Transcribe
Know the specialized APIs.
The exam often tests specific capabilities such as AnalyzeExpense, AnalyzeID, DetectPiiEntities, and Rekognition moderation.
Remember the zero-training advantage.
A managed AI service is usually the answer when the scenario says “with the LEAST development effort,” “fastest to implement,” or “no ML expertise available.”
Multi-service pipelines are common.
Many scenarios require Transcribe followed by Comprehend. Recognize this pattern.
Use custom features only when the standard service falls short.
For example, if the business needs to identify specialized objects, consider Rekognition Custom Labels. If the business needs custom categories in text, consider Comprehend custom classification.
Know the asynchronous vs. synchronous options.
Video analysis in Rekognition and large batch transcription in Transcribe are asynchronous. Image analysis and real-time streaming transcription are synchronous.

Traps

Do not use Rekognition for full document OCR.
Rekognition can detect text in images, but Textract is the service for extracting large amounts of text, tables, and forms from documents.
Do not use Textract for sentiment or entities.
Textract extracts text but does not understand it. Once text is extracted, Comprehend performs NLP.
Do not train a custom model for a solved problem.
If the exam says a standard invoice extraction task requires the LEAST effort, Textract is the answer. A custom SageMaker model or fine-tuned Bedrock model adds unnecessary work.
Do not use a foundation model for tasks that have a specialized managed service.
For example, do not use Amazon Bedrock to extract invoice totals when Textract AnalyzeExpense is purpose-built, faster, and cheaper.
Avoid mixing up audio and text.
Amazon Transcribe produces text from audio. Amazon Comprehend analyzes text. Amazon Polly produces speech from text, but it is not in this skill statement.
Watch for “real-time” keywords.
If the scenario says real-time transcription of live audio, use Transcribe streaming. If it says batch process stored audio, use batch transcription.
PII detection appears in multiple services.
Amazon Comprehend detects PII in text. Amazon Transcribe Call Analytics can redact PII in audio transcripts. Amazon Rekognition can detect faces in images. Choose based on the input type.

Summary Table

AWS AI Service	Input	Primary Output	Best Use Case
Amazon Textract	Scanned documents	Text, forms, tables, invoices, IDs	Accounts payable, KYC, medical forms
Amazon Rekognition	Images, video	Labels, faces, moderation, text	Content moderation, identity verification, security
Amazon Comprehend	Raw text	Entities, sentiment, key phrases, PII, topics	Customer support, social media monitoring, compliance
Amazon Transcribe	Audio	Transcripts, speaker labels, timestamps	Call analytics, meeting notes, subtitles

The key to this skill is not memorizing every API, but understanding which service solves which type of business problem and when a managed service is the correct approach compared to custom ML or foundation models.