Final mission: the portfolio ML system
Twelve chapters of machinery, one build. The capstone asks for the smallest complete system — not the biggest version of any one piece, but every subsystem present, wired to its neighbors, and sturdy enough to show a stranger.
The job: one project that chains chapters 35–47 end to end. Pick a dataset with a real prediction question (a public dataset you genuinely find interesting beats an impressive-sounding one you'll abandon), and build:
- Pipeline → dataset — staged ingest with contracts and a quality gate (ch.35–37), features with a leak-proof time cutoff (ch.36, 38).
- Models with receipts — the baseline ladder and a scoreboard (ch.39), evals with slices and thresholds (ch.41), every run tracked (ch.45).
- Ship it — registry with promotion (ch.45), a validated serving endpoint with health and version (ch.46), a CI-style gate script that would block a regression (ch.46).
- Keep it alive — structured prediction logs, one drift check, a written retrain policy (ch.47).
- Tell the story — the eight-section architecture doc (ch.47.5), including failure modes and the cost paragraph.
The starter is your acceptance test: a checklist that demands receipts, with an evidence string per subsystem — "built: True" without evidence is vibes, and you know better by now. The LLM-side capstone (chapter 25) plus this system are the two poles of the portfolio: one shows you can ship an AI feature; this one shows you can run a learning system. The doc is what makes either legible to a stranger — which is the actual test of engineering.