1…Nf6
+0.18, the engine’s eval after 1…Nf6. The seam rests there.
Result and evidence
- 0.87 AUC-ROCControlled evaluationOffline evaluation, no published comparator.
The case
- Problem
- Daily credit scores that a desk has to interpret, serve, and retrain, not a notebook that dies after the plot.
- Decision
- LightGBM and XGBoost, with SHAP explaining each score, served through BentoML. Kafka carries a daily retrain without taking the API down.
- Considered and rejected
- A single offline notebook.
Architecture
- Kafka
- LightGBM / XGBoost
- BentoML
Around it
- SHAPexplains each score
- Limitations
- Dataset size, split, leakage controls and prevalence were not recorded.
TensorFlow · XGBoost · LightGBM · BentoML · SHAP · Kafka
- The source is in a private repository.
- Back to all work