Machine Learning Engineering Principles

1. Reproducibility

• •Seeds: Always set random seeds for numpy, pandas, and the model framework (XGBoost/sklearn).
• •Config: Model hyperparameters must be decoupled from code (passed via args or config file), not hardcoded.

2. Artifact Management

• •Model Saving: Save trained models with versioning info.
  • •Path: models/{model_type}_{timestamp}.pkl
• •Metadata: Save a companion JSON with the model containing:
  • •Feature list (critical for inference alignment)
  • •Training date
  • •Performance metrics on validation set

3. Inference Pattern

• •Batch Inference: For this pipeline, prefer batch inference scripts (src/predict.py) that read from the Feature Store (DuckDB) and write predictions back to a Results table.
• •Validation: "Train-Serve Skew" checks. Ensure the distribution of features at inference time matches training time.