Short, practical write-ups of what I learned building things, with the ideas general enough to carry to the next project. Grouped as collections, one per project. Details generalized.
Five field notes on taking a chain of language-model agents from "a different answer every run" to reproducible: the harness, single-sourced facts, deterministic output, graded test data, and the drift eval that proves it holds.
Two notes from the referral analytics project, on the gap between certificate exercises and a real use case: first building a model you can trust (imbalance, Weight-of-Evidence encoding, the leakage a pipeline prevents), then deciding it's done and reading what it says.