Generative AI Developer Course: Azure AI-103 Live Training

The first domain is service selection and operational judgement, and it is where scenario questions are decided. Most candidates skip straight to prompting and lose marks here.

• Choosing between Microsoft Foundry, standalone Azure OpenAI and the discrete Azure AI services
• Model selection: capability, context window, latency and cost as a single decision
• Deployment options, quotas, scaling and provisioned throughput
• Security: managed identity, keyless credentials, private networking and role policies
• Responsible AI as architecture: filtering categories, severity levels and policy decisions
• Cost control, token accounting and capacity planning before anything ships
• Hands on: provision a Foundry project and deploy your first model securely

The largest domain on the exam starts here. This module is about making a model behave predictably enough that an application can depend on it.

• Chat completions, system design and prompt patterns that survive production
• Structured output and JSON schema enforcement for reliable downstream parsing
• Function calling and tool definitions done properly
• Streaming responses and the user experience patterns around them
• Prompt versioning, caching and batching for cost at volume
• Hands on: a generative feature inside a working application, not a notebook

Retrieval is where most candidates and most real projects fail. We teach the choices that decide quality rather than a single reference pipeline.

• Azure AI Search: indexes, indexers, skillsets and the knowledge store
• Embedding model choice and what actually changes when you switch
• Chunking strategies compared: fixed size, sentence boundary, semantic and parent child
• Hybrid search, semantic ranker and scoring profiles
• Citations, freshness and handling content the model must not invent around
• Measuring retrieval quality with a repeatable evaluation set
• Hands on: build a chunking comparison and prove which configuration wins

This is the material that separates AI-103 from every exam that came before it, and the reason the credential is worth holding early.

• Foundry Agent Service: threads, conversation memory and state
• Tool calling with function tools, OpenAPI tools and code interpreter
• Knowledge integration: choosing memory, tool and knowledge services per scenario
• Multi agent orchestration and handoff patterns between specialised agents
• Approval flows and human in the loop checkpoints for consequential actions
• Model Context Protocol and how agents reach external systems safely
• Hands on: an agent that completes a real multi step workflow end to end

An agent that acts without guardrails is a liability. The exam tests this domain directly and so does every enterprise procurement review.

• Content safety filters, categories and severity thresholds
• Prompt shields and jailbreak resistance
• Evaluators for groundedness, relevance, coherence, fluency and safety
• Trace logging and end to end observability across an agent run
• Red teaming your own system before somebody else does
• Hands on: an evaluation and safety harness wired into the agent from Module 4

Agents are no longer text only. This module covers the modalities the exam expects you to integrate as agent inputs and outputs.

• Image analysis and multimodal understanding for agent inputs
• Image and video generation workflows
• Speech to text and text to speech, real time against batch processing
• Speech as an agent modality, including custom speech models
• Translation with Azure Translator and with language model powered flows
• The Realtime API for low latency two way agents, and when to prefer the Speech SDK
• Hands on: give your agent a voice and a pair of eyes

The final domain plus a full exam rehearsal. Note that the exam may include interactive lab style components, so we practise pacing as well as content.

• Document Intelligence: prebuilt models, custom extraction and custom classification
• Content Understanding for clean, grounded representations that feed agents and RAG
• Multimodal extraction pipelines combining OCR, layout analysis and field extraction
• End to end skillset chains: OCR, language detection, entity recognition, embedding generation
• The full AI-103 outline reviewed in blueprint proportion
• A full length domain scored mock plus score report review and a revision plan

The Anthropic side of the program. You rebuild your agent on Claude, which is the fastest way to understand where each platform genuinely wins.

• Claude platform mastery: models, Projects, artifacts and tool use
• Building with the Claude API and with Claude Code
• Agentic patterns on Claude, including MCP servers and long running tool use
• Prompt caching, batching and cost control at production volume
• Anthropic certification preparation with practice questions and a mock
• Capstone: one product shipped on Foundry and rebuilt on Claude, with a written comparison and a mock interview