Week 12 Part V · Integration and AI-Enabled Systems

AI API Integration and Enterprise Best Practices

Instructor lesson plan: lecture (2 h) and practice (2 h).

Learning objectives

Integrate AI APIs through explicit application boundaries.
Use .NET AI libraries for chat, embeddings, tool calling, and orchestration.
Design prompts, context construction, policy checks, and evaluation as maintainable artifacts.

Tools this week

Microsoft.Extensions.AISemantic KernelOpenAI .NET SDKAzure.AI.OpenAI

🎓Lecture · 2 hours

0:00-0:20	20 min	AI as an enterprise subsystemUse cases where AI supports workflow without replacing architecture.
0:20-0:45	25 min	.NET AI library landscapeMicrosoft.Extensions.AI, Semantic Kernel, provider SDKs, and Azure integration.
0:45-1:10	25 min	Boundaries and adaptersTyped clients, prompt ownership, retries, rate limits, logging, and secrets.
1:10-1:20	10 min	Break
1:20-1:45	25 min	Embeddings and tool callingSemantic search, structured outputs, function calls, and safety constraints.
1:45-2:00	15 min	EvaluationTest sets, review samples, regression checks, and documentation of AI behavior.

Common misconception to confront.

Students often think: an AI call can sit directly in a controller.
Set it straight: AI calls need boundaries for retry, policy, logging, prompt versioning, and replacement.

Check for understanding

Why wrap an AI provider SDK?

To keep provider details, retries, policy checks, and prompt construction outside business workflows and controllers.

What should an AI evaluation include?

Representative inputs, expected properties, failure cases, and a repeatable way to compare behavior after changes.

Key takeaways.

AI features need ordinary engineering discipline.
Prompts and context are application artifacts.
Evaluation is part of integration design.

📚Reading & resources

Microsoft.Extensions.AI Provider-neutral .NET abstractions.
Semantic Kernel Orchestration, plugins, and function calling in C#.
OpenAI .NET SDK Official .NET SDK repository.

💻Practice · 2 hours

0:00-0:25	25 min	AI use-case selectionChoose one optional feature or document why AI is not appropriate for the project.
0:25-0:55	30 min	Client boundaryDefine an interface and adapter for an AI operation.
0:55-1:10	15 min	Prompt and response modelMake inputs, outputs, and failure cases explicit.
1:10-1:20	10 min	Break
1:20-1:50	30 min	Evaluation sketchCreate a small evaluation set and decide what counts as acceptable behavior.
1:50-2:00	10 min	Project-integration briefSubmit AI implementation or integration design.

Project integration (this week)

Add an optional AI-backed use case or document an AI integration design.
Keep AI calls behind a replaceable interface.
Add a short evaluation and risk note.