ML Optimization Lab by Mathurin Aché

Latest technical notes

• Model Monitoring Signals for Solo ML Products

  A reusable checklist for solo ML model monitoring with a frozen validation contract.

• Tabular Baseline in 30 Minutes

  A strict baseline recipe that gets from CSV to reliable validation quickly.

• Time-Series Cross-Validation Without Future Leakage

  A blocked CV template for demand, finance, and event datasets.

• XGBoost vs LightGBM vs CatBoost for High-Cardinality Features

  A decision table and benchmark notebook template for categorical-heavy data.

• Ranking Metrics Explained for Recommendation Models

  A practical guide to MAP@K, NDCG, Recall@K, and validation traps.

• Prompt Pack for Data Science Agents

  Prompts that force agents to produce tests, checks, and artifacts instead of vague advice.

• Out-of-Fold Ensembling Recipe for Tabular Competitions

  A minimal OOF stacking workflow with reproducible files.

• Public Dataset Teardown Workflow for ML Content

  A repeatable process to turn public datasets into useful articles and notebooks.

• Optuna Search Spaces for LightGBM on Tabular Data

  A Kaggle-style search-space recipe that prevents wasted trials.

• Leaderboard Overfitting Warning Signs

  A sanity checklist before trusting a public leaderboard jump.

• LightGBM Early Stopping and Seed Stability Checklist

  A reproducibility checklist for noisy leaderboard improvements.

• ML Interview Optimization Problems for Data Scientists

  A set of realistic prompts about metrics, CV, and model selection.