Course · Plus
Production AI on Azure
For .NET and C# developers moving into GenAI engineering on the Azure stack: Azure OpenAI, Azure AI Search RAG, keyless auth with managed identity, LangGraph agents, LangSmith and Langfuse tracing, OpenTelemetry, Bicep and Terraform, and SonarQube quality gates, ending with a secure, observable RAG capstone.
Open the courseChapters, a study plan and timed practice exams.Chapters
1. Azure OpenAI Service 60 min
Resources, deployments, endpoints and the v1 API, quotas, content filters and calling models from C# and Python.
- Explain the difference between a resource, a model and a deployment, and why your code passes a deployment name.
- Choose a deployment type (Global Standard, Data Zone, Standard, Provisioned, Batch) for a given latency, residency and cost need.
- Describe TPM and RPM quotas, what a 429 means and how to handle it.
- Explain the default content filters and what Prompt Shields add.
- Call a chat model from C# and Python with keyless (Entra ID) auth on the v1 endpoint.
2. Azure AI Search for RAG 70 min
Indexes, indexers and skillsets, vector and hybrid search, the semantic ranker, integrated vectorization and a full RAG pipeline.
- Design an index schema with searchable, filterable and vector fields.
- Explain indexers, data sources and skillsets, and what integrated vectorization automates.
- Compare keyword, vector, hybrid and hybrid plus semantic ranker queries.
- Build the retrieve-then-generate loop in C# or Python with citations.
- Explain why On Your Data was retired and what replaces it.
3. Managed identity, Entra ID and Key Vault 55 min
Keyless auth with managed identities and DefaultAzureCredential, least-privilege RBAC roles, and Key Vault for the secrets you cannot avoid.
- Explain system-assigned vs user-assigned managed identities and when to use each.
- Describe the DefaultAzureCredential chain and how the same code works locally and in Azure.
- Assign least-privilege data-plane roles for Azure OpenAI, AI Search, Storage and Key Vault.
- Turn off local (key) auth and keep secrets out of code, config and pipelines.
4. LangGraph: agents as state graphs 70 min
State, nodes, edges, conditional routing, cycles, checkpointing and human-in-the-loop, and when to pick LangGraph over a plain chain.
- Model an agent as a StateGraph with typed state, nodes and edges.
- Use conditional edges to route and to build tool-calling loops with a stop condition.
- Add a checkpointer for memory per thread, and pause for human approval with interrupt.
- Explain when a linear LangChain chain is enough and when you need a graph.
5. LangSmith and Langfuse 55 min
Tracing LLM apps as runs and spans, datasets and evaluations, prompt management, and choosing between LangSmith and Langfuse, cloud or self-hosted.
- Explain traces, runs, spans and generations, and what to capture on each LLM call.
- Instrument Python code with LangSmith @traceable and Langfuse @observe.
- Build a dataset and run an offline evaluation in CI.
- Compare LangSmith and Langfuse on licensing, hosting and integrations.
6. OpenTelemetry for AI apps 60 min
Traces, metrics and logs, OTLP and the Collector, the GenAI semantic conventions, and exporting .NET and Python apps to Azure Monitor.
- Explain traces, spans, metrics and logs and how context propagates between services.
- Describe OTLP, the OpenTelemetry Collector and why you would put one in the middle.
- Use the GenAI semantic conventions (gen_ai.* attributes and metrics) and know their stability status.
- Send .NET and Python telemetry to Application Insights with the Azure Monitor OpenTelemetry Distro.
7. Infrastructure as code with Bicep 60 min
ARM and Bicep, parameters and modules, what-if, and deploying an Azure OpenAI + AI Search + managed identity stack with role assignments.
- Explain how Bicep compiles to ARM templates and what Azure Resource Manager does with them.
- Write a Bicep file with parameters, variables, outputs and a module.
- Preview changes with what-if and deploy at resource group scope.
- Deploy Azure OpenAI with a model deployment, AI Search and a user-assigned identity with least-privilege roles.
8. Terraform on Azure 60 min
Providers, state and remote backends, plan and apply, modules, the azurerm provider, and when to choose Terraform or Bicep.
- Explain providers, resources, data sources, variables and outputs in HCL.
- Configure remote state in Azure Storage with locking and Entra ID auth.
- Use the init, plan and apply workflow in CI with a saved plan.
- Compare Terraform and Bicep and pick one for a given team.
9. SonarQube and quality gates 50 min
Static analysis of bugs, vulnerabilities, code smells and security hotspots, coverage, and failing CI when the quality gate fails.
- Describe what SonarQube reports: issues by quality (reliability, security, maintainability), security hotspots, coverage and duplication.
- Explain the clean-as-you-code idea and the default Sonar way quality gate on new code.
- Run analysis for a .NET solution with coverage in GitHub Actions or Azure DevOps.
- Make the pipeline fail when the quality gate fails.
10. Capstone: a secure, observable RAG app on Azure 3 to 5 days
Put it all together: keyless RAG over Azure AI Search and Azure OpenAI, traced with OpenTelemetry, deployed with Bicep or Terraform, behind a CI quality gate.
- Draw and explain an end-to-end production RAG architecture on Azure.
- Build and deploy it with IaC, managed identity and no secrets.
- Instrument it with OpenTelemetry and an LLM evaluation step.
- Present the design and its trade-offs in an interview.