Notes from production, not from a content calendar
What breaks in real AI systems, why it breaks and what we do about it. Written by the engineer doing the work, aimed at the people who have to make the build-or-buy decision and then live with it.
Why RAG systems fail in production, and the five fixes that matter
Teams swap the model, rewrite the prompt and raise the temperature. The answers stay wrong, because the correct passage was never in the context window to begin with.
When not to use a large language model
A generative model is the wrong tool at least as often as it is the right one. Knowing which situation you are in before the build starts is worth more than any prompt technique.
The checklist an AI agent has to pass before it runs unattended
A demo agent runs one happy path in front of an audience. A production agent runs thousands of times a week against APIs that rate-limit and records that contradict each other.
The writing is the same thing as the work
Every article here comes out of a system that shipped. When we describe a failure mode, it is one we were paid to fix, and when we recommend an approach it is one running in production somewhere with a client's name on it.
Want these by email?
There is no newsletter yet. The fastest way to follow the writing is LinkedIn, where these pieces go out first and where most of the technical discussion happens.
Reading this because something is not working?
Describe the system and where it is falling short. Thirty minutes is usually enough to identify whether the problem is retrieval, evaluation, architecture or scope.