18 articles
Notes on applied AI — RAG, MCP, agents — alongside architecture, API design, and code that stays readable long after it ships.
A small TypeScript server with one tool, tested with curl before any client touched it, and the three separate reasons Claude could not find it once it was installed.
Read articleThe admin app was sending tokens the API never checked. Closing that without breaking my teammates, opening exactly two endpoints for the MCP server, and moving the API to Azure Functions.
Read articleBinding to 0.0.0.0 switched off a protection without saying so, a health check could never pass, and what actually happens when you paste a URL into Claude.
Read articleWrite tools need to know who is calling. The client registration gap between Claude and Entra ID, why a platform subdomain cannot be the resource, and the on-behalf-of exchange that keeps the API honest.
Read articleThe design from part four, built. What Entra made me prove before it would accept the server as a name, why the client ID cannot live in my code, and testing the whole chain by hand before writing any of it.
Read articleThe architecture behind a conversational assistant that recommends events and articles from a real database: intent extraction, MongoDB vector search, and a grounded generation step.
Read articleBuild the RAG flow against local models, then run the same local server against Azure before deploying. Two phases, two purposes, and a pile of environment specific code nobody decided to write.
Read articleThe fixed three step chain has a ceiling, and I can see it from here. Where the routing decisions actually belong, why the tools want to be an MCP server, and the arithmetic that says do not get carried away.
Read articleA working definition of AI that survives contact with production, and why the shift from deterministic to probabilistic changes how you build.
Read articleEverything becomes numbers before a model can touch it. Understanding that conversion explains both the impressive capabilities and the bizarre failures.
Read articleModels do not make decisions, they calculate probabilities. Who sets the threshold, why confidence is not accuracy, and the arithmetic that breaks multi-step pipelines.
Read articleGarbage in, garbage out is too simple. The real failure is data that is clean but not representative, and the drift that starts the day you ship.
Read articleFour paths with real numbers attached, the vendor red flags worth listening for, and why starting from constraints beats starting from wishes.
Read articleTurning a naive API call into something that survives real traffic, queues, circuit breakers, the failure cliff, and the four metrics that tell you the truth.
Read articleA trained model is not a product. What happens between "the model works" and "the business depends on it", and how to keep it trustworthy once it is live.
Read articleWhy 95% of AI investments show no measurable return, what workslop is, and the cost models, FinOps thinking, and guardrails that keep AI usage sane at scale.
Read article
A practical comparison of REST and GraphQL with interactive examples
Read article
Essential principles for writing maintainable, readable, and scalable code
Read article