Writing
Notes on taking AI from a demo to production. Retrieval, evals, cost and latency, drawn from the work itself, not the brochure.
Why most RAG prototypes fail in production
A demo answers the questions it was built for. Production sends thousands, each with a cost, a latency and a way to fail. The gap is not the model. It is everything around it.
The numbers we check before an AI feature ships
Six measurements decide whether an AI feature is ready. Each needs a target set before launch, not a reading taken after. Here is the set we use and what each one catches.