AI Engineering FoundryLoading…
Loading…
Practical guides and engineering notes for building real-world AI systems — indexed by discipline.
Build an MCP server in C#: read-only SQL Server tools, a T-SQL parser gate, cost and row limits, and setup for Claude Desktop, VS Code and Claude Code.
Read article →Build a RAG API in C#: chunk by page and heading, embed with Ollama, search Qdrant with access filters, stream cited answers, and measure the results.
Read article →Run a local LLM with Ollama and call it from C#: HttpClient, OllamaSharp and IChatClient, streaming, JSON output, embeddings, plus troubleshooting.
Read article →What is RAG? Retrieval-Augmented Generation for developers: chunking, embeddings, vector search, prompts, evaluation, and failure modes, with C# examples.
Read article →Convert HTML to PDF in ASP.NET Core with SelectPdf's Blink engine: a .NET 8 minimal API that builds a 3-page A4 document with a StringBuilder, CSS print rules that control every page break, and a React front end that previews the HTML and downloads the PDF.
Read article →AI agent security for developers: why the prompt is never the security boundary, and how to handle prompt injection, tool permissions, unsafe model output, data leaks, authorization and runaway cost — with real failures and code from three .NET AI projects.
Read article →How to prevent SQL injection in AI-generated SQL: a least-privilege database user, a real T-SQL parser with an allow-list, cost limits, safe tool queries and a prompt that is never the security boundary. With code for ASP.NET Core and SQL Server.
Read article →ASP.NET Core vs Python FastAPI for AI APIs — the same streaming LLM endpoint built in both, and a decision framework based on where your model actually runs.
Read article →Build Text-to-SQL in C# for SQL Server: least-privilege execution, schema-aware prompts, ScriptDom validation and a repair loop. 53% to 88%, same model.
Read article →The baseline every Text-to-SQL demo starts from — table names in, SQL out — measured against 32 real questions. 53% correct, seven silently wrong answers, and an UPDATE and a DROP TABLE sent to the database.
Read article →Give the model what a human analyst would look at — types, keys, join paths, column descriptions, allowed values, business rules — as annotated DDL. Accuracy goes from 53% to 81%, and unsafe SQL executed drops to zero.
Read article →Treat model-generated SQL as untrusted input: parse it with Microsoft's ScriptDom, check it against an allow-list and the real schema, estimate its cost, and feed exact errors back for repair. 88% correct, zero failures, zero unsafe SQL.
Read article →Valid, safe SQL can still answer the wrong question. V4 retrieves a versioned business glossary from Qdrant — term match plus dense search — and lifts held-out accuracy from 50% to 75%.
Read article →Give the agent tools to inspect tables, search the glossary, sample values and test queries. Measured honestly: a pure agent scores 42% held-out, the model rarely calls a tool when it has context, and it rubber-stamps its own wrong answers.
Read article →OpenTelemetry traces and metrics across the whole pipeline, an audit log in a separate database, and a hardened container. Then a 132-question workload: the model is 96% of the time, and the first trace found a 2-second IPv6 localhost penalty on every new connection.
Read article →