Two kinds of post, kept apart on purpose. Findings are write-ups of systems I built, each one anchored to a run that went wrong and the measurement that caught it. Notes are explainers that stand on other people's work.
One system, one run that failed in a way nothing was watching for, and the measurement added afterwards. Every number is reproducible from the repository it came from.
Assay establishes whether a coding agent's repair actually solved the task. A hosted model, deepseek-v4-pro, spent 100 steps and 65 cents, finished with all 133 visible tests passing, and had fixed none of the eighteen that graded it.
Ratchet turns an observed integration gap into a scoped, evaluated change. Schema F1 rose to 0.91, the tests passed, a human approved it, and the mapping produced a value for zero of 1,931 records.
Vendor vs Valor builds a reviewable build-or-buy comparison without redoing the research every time the question comes back. It cites every source, argues with itself, and hands the call back to you.
Explainers on things I had to understand properly before I could use them. No original measurement; they stand on published work and say whose.
LLM latency isn't one number. It's prefill and decode, pulling against each other. A mental model for finding the milliseconds before you reach for a quantization library.
Most vector databases run the same core algorithm, so 'which is fastest' rarely decides anything. The differences that bite are operational: who runs it, how it filters, what it costs at scale.
Fine-tuning is the most over-reached-for tool in the LLM toolbox. A grounded guide to deciding between prompting, retrieval, and training, before you spend the weeks.