Notes on building AI systems that have to work in production — mostly the unglamorous parts. Written from the perspective of someone who spent 17 years on enterprise backends before the current wave.
When an agent writes nearly all the code, producing a change is cheap and knowing it is right is not. The verification pipeline I run around my coding agents — planning, independent QA, reviewers with a judge, a ledger for every finding — and the green builds that taught me to need each part of it.
Read →RAG is not one architecture, it is at least seven — and they fail in different ways. A field guide to the retrieval patterns that actually ship: what each one costs, what it fixes, and how the shape of your users' questions decides which one you need.
Read →I lead development of an agentic AI platform, which is exactly why I spend so much time talking people out of building one. A four-rung ladder from plain code to a full agent loop — and the one question that decides which rung your problem belongs on.
Read →Everyone's first agent is twenty lines: call the model, run the tool, append the result, repeat. What has to go in between those steps — validation, tool scoping, approvals, bounded execution, re-validation — before it survives real users.
Read →Before I worked on agents, I spent a year building business activity monitoring for a national telecom. Almost everything useful I know about debugging LLM systems, I learned there — and the industry is currently relearning it the hard way.
Read →