Generative AI for .NET Developers with Microsoft Foundry

AI hierarchy, supervised learning, unsupervised learning, deep learning, foundation models, generative AI, and agentic AI.

Tokens, tokenization, embeddings, transformer attention at a conceptual level, training versus inference, next-token prediction, and probabilistic generation.

Large and small language models, reasoning models, embedding models, multimodal models, image-generation models, speech models, and specialized extraction models.

System and user messages, context windows, temperature, output length, grounding, structured output, and why identical prompts may not produce identical responses.

Generative AI, text analysis, speech, vision, information extraction, search, RAG, classification, and summarization.

Fairness, reliability, safety, privacy, security, transparency, accountability, hallucinations, prompt injection, evaluation, monitoring, and governance.

Explore the Foundry control plane, project resources, models, agents, tools, evaluations, tracing, and application-development surfaces.

Provision the resource and project that will contain the VibeCast models, connections, tools, evaluations, and observability configuration.

Foundry Regions: https://learn.microsoft.com/azure/foundry/openai/how-to/responses#region-availability

Model suitability and routing require a real decision framework.

Compare model hosting through Microsoft Foundry, direct AI provider APIs, multi-model cloud platforms, and locally hosted OpenAI-compatible endpoints. Evaluate each option according to model choice, identity, data governance, latency, cost, portability, regional availability, and operational responsibility.

Recap of cloud infrastructure. Your endpoints are hot, secured, and ready for integration.

Moving from the cloud portal to your local integrated development environment (IDE).

Ensuring machine readiness by testing terminal runtimes to verify a stable modern installation target (.NET 10+).

Scaffolding a clean enterprise web application pipeline using the terminal CLI or Visual Studio templates.

Before writing code, we compare the developer surfaces available to .NET developers: Microsoft Foundry SDKs, OpenAI-compatible clients, Microsoft.Extensions.AI abstractions, and Microsoft Agent Framework.

• Select the application-facing abstractions, provider adapters, Foundry project client, service-specific clients, and test doubles used by the solution.

• Keep the broader SDK comparison in the Foundry architecture section.

Return incremental model output, propagate request cancellation, handle disconnected clients, and protect ASP.NET Core request lifetimes.

Conceptual check. Your local architecture is decoupled, abstracted, and thread-safe.

Separate concept generation from episode planning using a dedicated service with its own contracts, validation, tool calls, and persistence.

Persist the editorial brief and create a durable episode record containing the audience, objective, tone, language, and publishing details.

Open the episode workspace from its persisted ID, reload the editorial brief, and enforce owner-scoped access before AI planning begins.

Use IChatClient and Microsoft.Extensions.AI to request JSON Schema-based structured output that maps to a typed C# episode plan.

Generate a typed EpisodePlan, deserialize the structured response, and render segments, key messages, evidence needs, media, and risks.

Distinguish transient request failures from invalid AI output, then define one bounded repair attempt to control latency, cost, and behavior.

Move planner instructions into versioned prompt assets and track prompt and schema versions for testing, auditing, and reproducible output.

Validate the EpisodePlan with deterministic C# rules, send one targeted repair request, and accept or reject the repaired result.

Handle cancellation, timeout, refusal, invalid output, repair failure, and provider errors with explicit planning result states.

Use typed C# tool calling to provide the current application context while keeping authorization, validation, and execution under .NET control.

Create a typed editorial policy tool that returns duration limits, disclaimer rules, supported media, and publication constraints.

Register and invoke the editorial policy tool during planning, validate its arguments, and capture evidence of safe tool calls.

Persist an accepted EpisodePlan with prompt, schema, validation, repair, and tool metadata, then reload the versioned snapshot.

Review the complete workflow from persisted brief to structured generation, validation, bounded repair, typed tools, and saved episode plan.

Transitioning VibeCast past basic text into processing audio streams and visual generation fields.

Explain modality, multimodal requests, generated versus uploaded media, and why provider success does not equal application acceptance.

Explain how text instructions and image content form a single request, what the model can infer, the important limitations, and why descriptions remain proposals.

Use IChatClient with image content, return a typed visual result, validate it, render suggested alt text, and require explicit acceptance.

Explain text-to-image generation, prompt and output controls, safety filters, binary output, and the difference between generating and understanding images.

Generate one promotional image, decode it, validate its format and dimensions, preserve generation metadata, and display it in the episode workspace.

Validate a supplied audio file, generate a transcript, attach it to the episode, record locale/status, and render it in the workspace.

Process one realistic PDF, such as a sponsor brief, and extract mandatory messages, restrictions, dates, and deliverables into a typed contract.

Introducing data ingestion mechanics to ground models on dynamic corporate repositories.

Create provider-neutral embedding services, batch requests, validate dimensions, handle model changes, and manage embedding versions.

• Explain when you still need a direct Azure AI Search client for semantic ranking, platform-specific indexing, agentic retrieval, security, and advanced query control.

Define searchable, filterable, retrievable, sortable, and vector fields that support enterprise retrieval and source-level security.

Execute and compare vector, hybrid, filtered, and semantically ranked queries.

Restrict tool availability, authorize each operation, protect state-changing actions, and prevent excessive or arbitrary tool execution.

Define representative prompts, expected evidence, structured assertions, edge cases, safety tests, and regression scenarios.

Define evaluator selection, thresholds, confidence, judge-model limitations, and metric combinations.

Execute selected evaluators, inspect low-scoring cases, and record a regression decision.