Audit ML training workflows for wasted compute — missing early stopping, brute-force hyperparameter sweeps, absent mixed-precision, checkpoint gaps that force full restarts, redundant retraining, and idle accelerators. Use this skill whenever the user shares training scripts/configs (PyTorch, TensorFlow, JAX, HF Trainer), sweep configs (Optuna, W&B, Ray Tune), asks why training is slow or expensive, or wants a training sustainability review. Part of Lean Agentic AI Skills; emits lean-findings.json.
