Long-form tutorials and engineering notes on .NET, React, AI/ML, and SaaS architecture.
LLMs from first principles: tokenization, embeddings, RoPE, self-attention, multi-head, nanoGPT, pretraining, SFT, LoRA, QLoRA, DPO, vLLM serving.
The MLOps playbook: deploying models (REST/batch/streaming), versioning + registry + CI/CD, drift monitoring, cloud strategies, and a full end-to-end app.
How Mattrx swapped its Python ML service for ML.NET — regression, classification, and clustering trained in C#, in-process, no PhD required.
Full C# RAG chatbot: Semantic Kernel + Azure AI Search hybrid + reranker + streaming + tenant security. Mattrx Help: 4% hallucination, 91% recall.
Agentic vs Traditional RAG in .NET — when each wins, full Semantic Kernel code, the router that saved Mattrx $10,900/month on AI bills.
Prompt injection, data leakage, tenant breaches — seven real attacks on enterprise LLM apps, and the Mattrx production code that stops each one.
The finale — a year of running MCP in production at Mattrx: what held, what broke, what we'd do differently, and the whole series in one mental model.
The MCP servers you built aren't tied to one model. Here's wiring MCP into OpenAI (client-side + the hosted tool) and agent frameworks — one server, any model.
The protocol was never the enterprise blocker — identity was. Here's rolling MCP across a company: provision once, inherit on login, govern centrally.
An agent that 'feels off' tells you nothing. Here's how to make MCP observable — one trace per agent run, structured logs, and the metrics that matter.
A tool that takes minutes shouldn't freeze your agent. Here's streaming, progress notifications, and the async-job pattern for long-running MCP tools.
A tool result can carry the attack. Here's how to secure MCP against prompt injection, poisoned servers, and tool abuse — and break the lethal trifecta.