DEV Community

Brent G Saucedo
Brent G Saucedo

Posted on

Azure AI-102 Azure AI Engineer Exam Cheat Sheet (Updated 2026)

#azure #ai #machinelearning #certification

The Designing and Implementing a Microsoft Azure AI Solution (AI-102) exam is for developers who build AI into apps. It is not a data science exam (that's DP-100). It is about APIs, SDKs, architecture, and security.

If you are preparing for AI-102, this cheat sheet covers the critical services, decision matrices, and implementation details you need to know.

1. AI Architecture & Security

Before you code, you must secure the resource.

Resource Types

  • Multi-Service Resource: Access to Vision, Language, Search, etc., with a single key/endpoint. Good for development/prototyping.
  • Single-Service Resource: Access to only one service (e.g., "Computer Vision" resource). Required if you need fine-grained cost tracking or specific tiers (e.g., Free Tier F0).

Authentication

  • Subscription Keys: Easiest, but least secure. Rotating keys requires app downtime unless you use Key Vault.
  • Managed Identity: The "Best Practice." No keys in code. Assign an RBAC role (e.g., Cognitive Services User) to the VM/App Service.
  • Key Vault: Store keys here and access via Secret URI.

Containers (Docker)

  • Use Case: Compliance (data cannot leave premises) or low-latency edge computing.
  • Billing: Containers are not free. They must connect to the Azure Billing Endpoint every 15 minutes to report usage.

2. Computer Vision & Custom Vision

Knowing which service to pick is 50% of the battle.

The Decision Matrix

Service Use Case Example
Computer Vision Pre-trained models. Generic analysis. "Describe this image." "Is there a dog?" "Generate a thumbnail."
Custom Vision You bring your own training data. Specific domains. "Is this my specific product?" "Is this screw defective?"
Face API Human face analysis. Identity verification, emotion detection, age estimation.

Custom Vision Types

  1. Classification: "What is this image?" (Output: Tag Toaster 98%).
    • Multiclass: One tag per image (Dog OR Cat).
    • Multilabel: Multiple tags per image (Dog AND Grass AND Ball).
  2. Object Detection: "Where is the object?" (Output: Bounding Box coordinates [x,y,w,h] + Label).

The "Read" API (OCR)

Standard OCR is synchronous. The Read API is asynchronous for large documents (PDFs/Images).

  1. POST request to analyze endpoint.
  2. Receive 202 Accepted + Operation-Location header.
  3. GET request to Operation-Location loop until status is succeeded.

3. Natural Language Processing (Language Service)

The consolidation of LUIS and Text Analytics.

Conversational Language Understanding (CLU)

Replaces LUIS.

  • Intent: What the user wants to do (e.g., BookFlight).
  • Entity: The parameters of the action (e.g., Paris, Tomorrow).
  • Utterance: What the user actually said (e.g., "Fly me to Paris tomorrow").
  • None Intent: Crucial for handling garbage input. Always required.

Key Capabilities

  • Sentiment Analysis: Returns confidence scores (0 to 1) for Positive, Neutral, Negative.
  • Key Phrase Extraction: Pulls main points ("The food was good but the service was slow" -> "food", "service").
  • Entity Linking: Connects words to Wikipedia/Knowledge Graph (e.g., "Venus" -> Planet vs. "Venus" -> Goddess).

Translator Service

  • Translate: Text to Text.
  • Transliterate: Convert script (e.g., Japanese characters to Latin alphabet) without translating the meaning.
  • Profanity Filtering: Settings: NoAction, Marked (***), or Deleted.

4. Knowledge Mining (Azure AI Search)

Formerly Azure Search. The most complex topic on the exam.

The Enrichment Pipeline

  1. Data Source: Where data lives (SQL, Blob, Cosmos).
  2. Indexer: The engine that crawls the data.
  3. Skillset: The AI processing steps (OCR, Translate, Entity Extraction).
    • Built-in Skills: Microsoft provided.
    • Custom Skills: Call a Function App/Web API for custom logic.
  4. Index: The searchable JSON document store.

Key Concepts

  • Push vs. Pull:
    • Pull: Indexer crawls data (SQL/Blob).
    • Push: Your code pushes JSON directly to the index (good for rare data sources).
  • Knowledge Store: Saves the enriched data (e.g., the text extracted from images) into tables/blobs for other apps to use (like PowerBI).

Search Syntax

  • Full Text: search=run (Finds "run", "running", "runner").
  • OData Filter: $filter=category eq 'Luxury' (Exact match).

5. Document Intelligence (Form Recognizer)

Extracting Key-Value pairs from documents.

Model Types

  • Read: Just text extraction (OCR).
  • General Document: Pre-trained for common structures.
  • Prebuilt: Invoices, Receipts, ID Cards, Tax Forms (W2).
  • Custom Template: You label data based on visual layout.
    • Requirements: Minimum 5 examples of the same layout to train.
  • Custom Neural: Understands complex, unstructured documents.
    • Requirements: Slower to train, more expensive.

6. Conversational AI (Bot Framework)

Building the interface.

  • Bot Framework Composer: Visual drag-and-drop tool to build dialogs.
  • Adaptive Cards: JSON snippets that render UI (Buttons, Forms, Images) consistently across any channel (Teams, Web, Slack).
  • Direct Line: The channel used to connect a bot to a custom mobile app or website.

📝 Exam "Gotchas"

  1. Content Safety: Know the difference between breaking (error) and flagging (warning) content.
  2. Video Indexer: It provides insights like OCR in video, Face detection, and Speaker diarization (Who spoke when?).
  3. Private Endpoints: If a question mentions "Banking" or "High Security," the answer almost always involves disabling public access and using Private Link.
  4. Speech Translation: You can translate Speech-to-Text (See words on screen) or Speech-to-Speech (Hear translated audio).
  5. Region Support: Not all features (especially Neural Voice) are available in every Azure region.

Good luck with your AI-102!

Top comments (0)