ML 工程评审官

# MLE Reviewer (ML 工程评审官) ## Overview **MLE Reviewer (ML 工程评审官)** is a senior machine-learning engineering reviewer who moves model code from "works in a notebook" to production-safe ML systems. It reviews data contracts, feature pipelines, training reproducibility, offline/online evaluation, model serving, monitoring, and rollback — surfacing leakage, irreproducible promotion, unsafe serving, and missing observability before they ship. ## Key Capabilities - ✅ **Leakage prevention** — point-in-time joins, split discipline, no future labels - ✅ **Reproducibility** — code/config/seed-driven training, versioned artifacts - ✅ **Promotion gates** — fail-closed, baseline-compared, slice-aware evaluation - ✅ **Serving safety** — schema rejection, timeouts, fallback, shadow/canary/A-B rollback - ✅ **Monitoring & response** — drift, label arrival, alert ownership, runbooks - ✅ **Review-lane composition** — delegates security/code/performance to matching reviewers ## Usage 1. Load the `mle-reviewer` agent into your workspace. 2. @mention `@mle-reviewer` (or route via workflow) on ML/MLOps, training, inference, feature-store, or evaluation code changes. 3. It inspects the diff, runs lightweight checks, applies the production-ML checklist, and returns a severity-ordered report ending with `APPROVE | APPROVE WITH WARNINGS | BLOCK`. ## Licensing Source Based on the **mle-reviewer** role definition from [everything-claude-code](https://github.com/affaan-m/everything-claude-code) (MIT). --- **Attribution:** Based on [mle-reviewer agent](https://github.com/affaan-m/everything-claude-code/blob/main/agents/mle-reviewer.md) from [everything-claude-code](https://github.com/affaan-m/everything-claude-code) - Original author: Affaan Mustafa - License: MIT