Focus
Models
Fine-tuning, and the gates that decide whether a fine-tune is allowed to ship.
I treat training as a last-mile tool. Most product failures I see are retrieval, evaluation, or orchestration — not a missing LoRA. So the interesting artifact usually isn't the weights; it's the gate that tells you whether the weights earned their place.
Nothing on this shelf has shipped yet. The design below is worked through on paper, not built; the two notes are about when training is — and isn't — the right call. When there's a real fine-tune with a real gate behind it, it goes here.
On this shelf
Fine-tuning is a last resort, not a first moveNotesWhen training helps, when it hides a data problem, and how I decide between prompting, RAG, and weights.OpenSmall-model behavior does not predict large-model behaviorNotesWhy a prototype on a 7B can lie about production on a frontier model — and what to measure instead.OpenDoRA × KAN: nonlinear weight adaptersDesigned, not builtA PEFT method that replaces the linear low-rank adapter with a B-spline KAN on the directional component. Math worked through; no implementation and no repository.Not built