Production-workflow для ML-инженерии

★ 7.9 · ai-tooling

mle-workflow is a Claude Code skill that brings production machine-learning engineering practices to model development: data contracts, reproducible training, measurable quality gates, deployment, monitoring, and rollback. It guides the conversion of notebook code into reusable training, evaluation, and inference pipelines suitable for ranking, search, recommendations, classifiers, forecasting, embeddings, anomaly detection, and LLM workflows. The skill deliberately avoids imposing a single architecture — it adapts to the data stores, serving mode, and rollout mechanics already in use on a project. It integrates with python-testing, pytorch-patterns, eval-harness, deployment-patterns, and docker-patterns. Teams reach for it when debugging data drift, label leakage, or stale-artifact failures, and when adding canary rollouts, shadow traffic, or post-deploy quality checks.