ML Research Engineering

Staged reproduction from inference-first to full retrain with deviation journals.

One-factor-at-a-time with matched budgets separating real gains from tuning luck.

Every run reproducible from logged config + code version + data snapshot reference.

Benchmark suites matched to target capabilities with contamination checks and honest scopes.

Provenance, composition, collection process, and limitations recorded before modeling begins.

Drafter selection, acceptance-rate monitoring, and batch-interaction effects in production serving.

Compute budgets, data-wall caveats, and extrapolation ranges separating signal from slide-ware.

Scaling-law claims converted into budget-honest training choices.