# Machine Learning Domain

## Approach
- Reproducible: pin deps, seeds, data versions.
- Separate data prep, training, evaluation, serving.
- Track experiments (MLflow/W&B).
- Validate data quality before training.
- Serve behind an API with monitoring.

## Common Patterns
- Train/val/test split with stratification.
- Cross-validation.
- Feature pipelines persisted with model.
- Batch + online inference.

## Testing
- Unit: preprocessing, feature extraction.
- Integration: training smoke test.
- Model: metric thresholds on holdout.

## Pitfalls
- Data leakage between splits.
- Non-reproducible results.
- Overfitting without validation.
- Serving stale models.
- Training/serving skew.
"