Preparing for AI-102: The Azure AI Engineer Deep Dive

Part 2 of the "From .NET to AI Engineer" series

From Fundamentals to Engineering

If AI-900 was learning to read sheet music, AI-102 is learning to conduct the orchestra.

The AI-102: Azure AI Engineer Associate certification validates your ability to design and implement AI solutions using Azure AI services. This isn't about understanding concepts—it's about building production systems.

As a .NET developer who's been integrating AI services for years, I thought this would be straightforward. I was humbled. There's a significant gap between "calling an API" and "architecting an AI solution."

What AI-102 Covers (2025/2026 Updated)

Microsoft refreshed the exam in 2025 to emphasize generative AI and agentic solutions. Here's the breakdown:

1. Plan and Manage an Azure AI Solution (15-20%)

This is the "architect" section. Before writing code, you need to:

Select appropriate Azure AI services

Use Case → Service Mapping:

"Extract text from scanned documents" → Azure AI Document Intelligence
"Build a custom chatbot" → Azure OpenAI + Azure AI Search
"Detect objects in images" → Azure AI Vision (Object Detection)
"Translate documents" → Azure AI Translator
"Analyze customer sentiment" → Azure AI Language
"Build custom ML models" → Azure Machine Learning

Plan for responsible AI

This goes deeper than AI-900. You need to understand:

• Content filtering implementation in Azure OpenAI
• Data privacy in AI pipelines
• Model monitoring and drift detection
• When to use human-in-the-loop systems

Manage costs and quotas

Real-world concern: Azure OpenAI tokens aren't free. You need to understand:

• Token counting and estimation
• Rate limiting strategies
• Provisioned throughput vs. pay-as-you-go

2. Implement Computer Vision Solutions (15-20%)

Azure AI Vision has evolved significantly. Key capabilities:

Image Analysis 4.0

// Modern approach using Azure AI Vision SDK
var client = new ImageAnalysisClient(
    new Uri(endpoint),
    new AzureKeyCredential(key));

var result = await client.AnalyzeAsync(
    imageUrl,
    VisualFeatures.Caption | VisualFeatures.Tags | VisualFeatures.Objects);

Console.WriteLine($"Caption: {result.Value.Caption.Text}");
// Output: "A person cooking in a modern kitchen"

Custom Vision vs. Vision Studio

• Custom Vision: Train your own image classification/object detection models
• Vision Studio: No-code interface for testing and prototyping

Face API Limitations

Microsoft has restricted facial recognition capabilities. You can detect faces and attributes, but identification requires approval for limited use cases. Know these restrictions for the exam.

3. Implement Natural Language Processing Solutions (20-25%)

This is the largest section—NLP is everywhere in enterprise AI.

Azure AI Language Features

Feature	Use Case
Sentiment Analysis	Customer feedback analysis
Key Phrase Extraction	Document summarization
Named Entity Recognition	Extract people, places, organizations
PII Detection	Compliance and data protection
Custom Text Classification	Domain-specific categorization
Conversational Language Understanding (CLU)	Intent recognition for bots

Building a CLU Model

Conversational Language Understanding replaced LUIS. The workflow:

1. Define intents (what users want to do)
2. Define entities (key information to extract)
3. Provide utterance examples
4. Train and evaluate
5. Deploy to a prediction endpoint

Intent: BookFlight
Entities: Origin, Destination, Date

Training utterances:
- "Book a flight from Toronto to Vancouver on March 15"
- "I need to fly from Edmonton to Montreal next Tuesday"
- "Can you get me a ticket to Calgary?"

4. Implement Knowledge Mining Solutions (10-15%)

Azure AI Search (formerly Cognitive Search) is powerful but complex.

The Knowledge Mining Pipeline

Data Sources → Indexer → Skillset → Index → Search
   ↓              ↓          ↓         ↓        ↓
Blob/SQL    Scheduling   AI Skills  Schema   Queries

Built-in Cognitive Skills

• OCR (extract text from images)
• Key phrase extraction
• Entity recognition
• Language detection
• Sentiment analysis
• Image analysis

Custom Skills

You can add Azure Functions as custom skills in the pipeline. This is how you integrate your own logic.

5. Implement Document Intelligence Solutions (10-15%)

Azure AI Document Intelligence (formerly Form Recognizer) extracts structured data from documents.

Prebuilt Models

• Invoices
• Receipts
• Business cards
• ID documents
• W-2 tax forms

Custom Models

Train models on your own document types. The exam tests:

• When to use prebuilt vs. custom
• Labeling strategies for training data
• Model composition (combining multiple models)

6. Implement Generative AI Solutions (10-15%)

This section was significantly expanded in 2025.

Azure OpenAI Service

var client = new OpenAIClient(
    new Uri(endpoint),
    new AzureKeyCredential(key));

var chatCompletionsOptions = new ChatCompletionsOptions
{
    DeploymentName = "gpt-4o",
    Messages =
    {
        new ChatRequestSystemMessage("You are a helpful assistant."),
        new ChatRequestUserMessage("Explain Azure AI services in one paragraph.")
    },
    MaxTokens = 500,
    Temperature = 0.7f
};

var response = await client.GetChatCompletionsAsync(chatCompletionsOptions);

Retrieval-Augmented Generation (RAG)

The most important pattern for enterprise GenAI:

1. User asks a question
2. Search relevant documents (Azure AI Search)
3. Include retrieved context in the prompt
4. LLM generates grounded response

User: "What's our refund policy?"

Without RAG: GPT makes something up
With RAG: Search finds policy document → GPT quotes actual policy

Prompt Engineering for the Exam

Know these patterns:

• Zero-shot: Direct question, no examples
• Few-shot: Provide examples before the question
• Chain-of-thought: "Let's think step by step..."
• System prompts: Setting context and constraints

My Study Plan

Timeline: 6-8 weeks with hands-on labs

Week 1-2: Azure AI Vision + Document Intelligence

• Complete MS Learn modules
• Build a receipt scanner using Document Intelligence
• Create a custom vision model for a personal project

Week 3-4: Natural Language Processing

• Deep dive into Azure AI Language
• Build a CLU model for a sample chatbot
• Implement sentiment analysis on real data

Week 5-6: Azure AI Search + Knowledge Mining

• Set up a search index with cognitive skills
• Build a custom skill using Azure Functions
• Implement semantic search with vectors

Week 7-8: Azure OpenAI + Generative AI

• Build a RAG solution end-to-end
• Implement content filtering
• Practice prompt engineering patterns

Resources

Official:

• AI-102 Study Guide
• AI-102T00 Training Course
• Exam Readiness Zone Videos

Hands-on:

• Azure AI Services Quickstarts
• GitHub: Microsoft AI-102 Labs repository

Questions That Clarify the "Why"

These questions helped me understand AI-102 concepts at a deeper level:

How would you build a document processing pipeline? → Azure AI Document Intelligence + Azure AI Search + Logic Apps/Functions

How do you prevent hallucinations in GPT responses? → RAG pattern, grounding, content filtering, temperature tuning

What's the difference between Azure OpenAI and OpenAI directly? → Enterprise security, compliance, private networking, SLA guarantees

How would you handle PII in AI workloads? → Azure AI Language PII detection, data masking, compliance controls

---

Key Takeaways

• AI-102 is significantly harder than AI-900—expect 6-8 weeks of serious study
• Hands-on labs are essential; theory alone won't pass this exam
• The 2025 refresh added significant generative AI and RAG content
• Focus on the "when to use what" decision-making, not just how things work

---

Next up: I'll explore the MCP (Model Context Protocol) revolution and what it means for .NET developers building AI agents.