Writing
Practical notes from the field
LLM engineering, evals, retrieval, and data quality for AI - written the way I would explain it to a team mid-build. Also available viaRSS.
November 11, 2026November 11, 2026
Evals are unit tests for non-deterministic code
If a prompt change can silently regress your product, you do not have an engineering process - you have a demo. A practical starting point for LLM evaluation suites.
[ evals ] [ LLMOps ]
Most LLM projects fail on the data layer, not the model layer
Teams debug prompts when they should be auditing retrieval. The failure mode I see most in stalled LLM projects is a data problem wearing a model costume.
[ LLM engineering ] [ data quality ]
Next step
Want to argue with me about this?
If a post above is wrong about something, or you are facing the problem it describes, I would genuinely like to hear about it.
Book a callUsually replies within 1-2 business days