mle · experimental

Context skill for the full ML engineering lifecycle: research, data pipelines, distributed training, evaluation, observability, and model publishing.

Philosophy

ML engineering is a systems problem, not just a modeling problem. A model that trains but can't be reproduced, monitored, or deployed is an experiment, not an asset. These recipes treat the entire lifecycle — from data ingestion to production serving — as an engineering system with the same rigour applied to any other distributed system: versioning, observability, regression prevention, and fault tolerance.

Recipes

• Researching a New Problem Domain — literature review, dataset discovery, baseline establishment, problem framing
• Data Pipelines with Ray Data — distributed loading, transformation, feature engineering, streaming to training
• Storage Formats for ML — Parquet, Arrow, HDF5, LMDB; when to use each; dataset versioning
• Distributed Training with Ray Train — DDP with Ray, Lightning integration, fault tolerance
• Model Training and Evaluation — training loop patterns, validation strategy, metrics, early stopping
• Observability with TensorBoard and Lightning — logging metrics, gradients, images; comparing runs; profiling
• Experiment Tracking with MLflow — runs, model registry, regression prevention, CI integration
• Model Publishing — ONNX, TorchScript, MLflow registry, Ray Serve, FastAPI