Precision in the Catalog

Success on the AWS Certified AI Practitioner exam depends on your ability to map a natural language requirement to a specific AWS technical service. Often, two services will seem similar.

Is it Comprehend or Rekognition? Is it Bedrock or SageMaker JumpStart?

In our final lesson of Module 4, we will build a "Decision Tree" to ensure you never pick the wrong service again.

---

1. The "Keyword" Trigger Map

AWS uses specific keywords in its questions to signal the correct answer.

If you see this keyword...	Use this service...
"Scanned PDF", "Form data", "Table extraction"	Amazon Textract
"Sentiment", "PiI Redaction", "Key phrases"	Amazon Comprehend
"Face detection", "Object moderation", "Landmarks"	Amazon Rekognition
"Natural speech", "Voice for a bot"	Amazon Polly (Text-to-Speech)
"Subtitle generation", "Audio to Text"	Amazon Transcribe (Speech-to-Text)
"Pre-trained Foundation Models", "Serverless GenAI"	Amazon Bedrock
"Custom ML Lifecycle", "Ground Truth", "Jupyter"	Amazon SageMaker

---

2. Decision Tree: Analyzing Images and Video

graph TD
    A[Requirement: Process Visual Data] --> B{Is it a document/PDF?}
    B -->|Yes| C{Do you need to extract TABLES?}
    C -->|Yes| D[Amazon Textract]
    C -->|No: Just read text| E[Amazon Rekognition or Textract]
    
    B -->|No: It's a photo/video| F{Need to check for 'Safety'?}
    F -->|Yes| G[Amazon Rekognition: Moderation]
    F -->|No| H{Need to recognize specific people?}
    H -->|Yes| I[Amazon Rekognition: Faces/Celebrities]

---

3. Decision Tree: Analyzing Language and Text

graph TD
    A[Requirement: Process Text/Language] --> B{Do you need to FIND things?}
    B -->|Yes| C{Is it 'Entities' or 'Sentiment'?}
    C -->|Yes| D[Amazon Comprehend]
    
    B -->|No: I need to CREATE text| E{Use a Foundation Model?}
    E -->|Yes| F[Amazon Bedrock]
    
    A -->|Requirement: Process Speech| G{Audio to Text?}
    G -->|Yes| H[Amazon Transcribe]
    G -->|No: Text to Audio| I[Amazon Polly]

---

4. The "SageMaker" Exception

There is one rule that overrides everything: Domain Specificity.

If an exam question says: "A company needs to detect defects on a specialized high-speed silicon manufacturing line that involves ultraviolet imaging data never seen before by standard vision systems," the answer should be Amazon SageMaker.

Why? Because Amazon Rekognition is trained on "Standard" images. It doesn't know what a "Silicon defect in ultraviolet light" looks like!

---

5. Summary: The Selection Heuristic

1. Step 1: Identify the Medium (Vision, Text, Audio).
2. Step 2: Identify the Task (Classify, Extract, Create).
3. Step 3: Check for Constraints (Time, Skill level, Domain uniqueness).
4. Step 4: Select the Managed Service unless the data is too unique.

---

Exercise: The Multi-Step Service

A travel agency wants to:

1. Allow customers to upload a photo of their passport.
2. Automatically pull the Name and Date of Birth from the passport.
3. Detect if the customer's face in the passport photo matches their current webcam photo.

Which two services should they use?

Answer:

1. Amazon Textract (to pull structured Name/DOB data from the document).
2. Amazon Rekognition (to perform the Face Comparison between the two photos).

---

Recap of Module 4

We have navigated the AWS landscape:

• We distinguished between AI (Managed), ML (Custom), and Analytics (Foundational Data).
• We explored the Managed vs. Custom trade-off.
• We built Decision Trees for service selection.

---

Knowledge Check

---

What's Next?

We’ve learned the landscape; now it’s time to move into the "Kingdom of AI Services" in detail. In Module 5: AWS AI Services (Pre-Trained), we will take a deep dive into Rekognition, Comprehend, and the rest of the gang.