AI-103: Developing AI Apps and Agents on Azure for Python Developers

Deploy three Foundry models, then benchmark them on latency, token usage, and tool-calling support using the model-explorer starter project. You'll finish compare_models.py and toolcall_probe.py using the Responses API via AIProjectClient.get_openai_client().

Finish the Bicep template for a full Foundry stack (hub, project, AI Search, Storage, App Insights), deploy it with a shell script, then add a GitHub Actions workflow that redeploys on push using OIDC federated credentials.

Migrate a FastAPI chat endpoint from AZURE_OPENAI_API_KEY to keyless authentication using DefaultAzureCredential and AIProjectClient.get_openai_client(), then deploy bicep/identity.bicep to provision a user-assigned managed identity with the Azure AI User role assignment on the Foundry scope.

Hands-on lab for AI-103 Lesson 5. You will build a responsible-AI orchestrator using Azure AI Content Safety as a pre/post filter, run Foundry Evaluations (groundedness, relevance, violence, hate) against an eval dataset, and add a human approval gate. Starter repo: ai-103-responsible-ai.

Hands-on lab for AI-103 Lesson 6. You will implement a FastAPI /chat endpoint that calls the OpenAI Responses API via AIProjectClient.get_openai_client(), wire three function tools (get_weather, calculate, lookup_inventory), chain tool outputs with previous_response_id, and add streaming. Starter repo: ai-103-chat-tools.

Note: This lab pre-provisions Azure resources at start — allow up to 10 minutes for the environment to become ready before beginning the exercises.

In this hands-on lab you fine-tune a gpt-4.1-mini foundation model on Summitline Outfitters' curated gear-recommendation dataset, deploy the fine-tuned model on a Developer-SKU deployment, and compare it against the base model on held-out Summitline test prompts. You will inspect the JSONL training format, submit a Supervised Fine-Tuning (SFT) job against the Foundry Azure OpenAI endpoint, poll job status + inspect loss / validation metrics, deploy the trained artifact keylessly, and eyeball a side-by-side quality comparison that shows the fine-tune consistently names in-catalog SKUs (pattern NW-SL-###) and cites the tool it used — where the base model does not.

Hands-on lab for AI-103 Lesson 7. You will create an Azure AI Search index, ingest and embed documents from sample_docs/, implement a /chat endpoint grounded on hybrid + semantic search, and run azure-ai-evaluation groundedness + relevance evaluators against eval_data.jsonl. Starter repo: ai-103-rag-eval.

Build a "Summitline concierge" agent with three function tools (get_weather, calculate, lookup_inventory) using the Azure AI Agents Service. You will author the AgentsClient + ToolSet wiring, drive a multi-turn thread, handle automatic tool execution, and pull a transcript back out.

Attach three knowledge tools to a Summitline Outfitters agent: FileSearchTool backed by a vector store of product PDFs, AzureAISearchTool bound to an AI Search connection, and a custom Content Understanding invoice extractor. Test routing by sending catalog, KB, and invoice questions at the same agent and inspecting citations.

Stand up an orchestrator agent that delegates to refund + lookup worker agents via ConnectedAgentTool, add a FastAPI approval queue that pauses refunds over $100 until a human approves, and wire OpenTelemetry through azure-monitor-opentelemetry so every span lands in Application Insights.

Hands-on companion for Lesson 11. Students complete four Vision exercises in the ai-103-vision FastAPI starter: generate, edit/inpaint, caption, and caption+detect with drawn bounding boxes. Keyless auth, Azure Validator checks each endpoint, and a capstone AI assessment closes the lab.

Note: This lab pre-provisions Azure resources at start — allow up to 10 minutes for the environment to become ready before beginning the exercises.

You are on the Summitline Outfitters engineering team building the video-analytics side of the product-support toolset. In this hands-on lab you will (1) upload sample product-demo videos to Azure Blob Storage, (2) author a Content Understanding pro-mode video analyzer with a Summitline-specific JSON schema (products demonstrated, key features shown, scene summary, presenter quality, duration), (3) invoke the analyzer and inspect the segmented output, (4) index each analyzer segment into Azure AI Search with vector embeddings from text-embedding-3-large, (5) wrap the whole flow as a Foundry agent FunctionTool that returns results with timestamped citations, and (6) run an end-to-end query — "show me videos where the tent is set up in windy conditions" — that reasons across the indexed segments. All auth is keyless via DefaultAzureCredential; no API keys are used.

Hands-on companion for Lesson 12. Students complete three exercises in ai-103-text-speech: Azure AI Language analyze (entities+sentiment+opinions), Speech STT/TTS round-trip, and Translator v3.0 text translation. Azure Validator confirms each FastAPI endpoint.

Note: This lab pre-provisions Azure resources at start — allow up to 10 minutes for the environment to become ready before beginning the exercises.

Build the hands-free product-question voice agent for Summitline Outfitters using the Azure Voice Live API and a Microsoft Foundry hosted agent. You will deploy a Foundry agent backed by an Azure AI Search RAG index over Summitline product docs (tent, backpack, sleeping bag, jacket, boots), wire the agent behind a Voice Live real-time WebSocket session, drive it with pre-recorded WAV files (the container is headless — no microphone), demonstrate barge-in interruption, and register a check_inventory(sku) function tool the agent can invoke mid-conversation. All auth is keyless via DefaultAzureCredential.

Hands-on companion for Lesson 13. Four exercises in ai-103-ingest-extract pipeline: Content Understanding extract with analyzer lifecycle and polling, chunk+embed+ingest, create HNSW+semantic index, hybrid semantic query with captions and answers.