Home / Notes
Written from systems I run in production, with the failures included. If something here reads like a warning, it is because I walked into it first.
Anyone can build the demo in an afternoon. What decides whether the project ships is permissions, knowing which version of each document is current, and citing the source every time. What stretches these projects is almost never the AI.
For months my clinical platform displayed a confidence score that did not come from the model we thought it did. Why it happens so often, how to check yours in an afternoon, and what to show instead.