Автоматизация ML-экспериментов с деплоем на GPU

★ 7.1 · engineering

experiment-bridge is a Claude Code skill that bridges idea discovery and the auto-review loop by turning a structured experiment plan into running GPU jobs with initial results. It reads `EXPERIMENT_PLAN.md`, `EXPERIMENT_TRACKER.md`, and `FINAL_PROPOSAL.md`, then implements training and evaluation scripts with argparse, fixed seeds, and JSON/CSV result logging before deploying via `/run-experiment`. When `CODE_REVIEW` is enabled, the generated code is reviewed by GPT-5.6-Sol at xhigh reasoning effort to catch logic bugs before spending GPU hours. Configurable flags include `SANITY_FIRST` (run the smallest experiment first), `AUTO_DEPLOY`, `MAX_PARALLEL_RUNS` (up to 4 parallel runs), and `BASE_REPO` for cloning an existing codebase as the starting point. The skill targets ML researchers who need to automate the full pipeline from a written experiment plan to collected baseline results.