How rajdeepofgithub/sbdr works
SBDR is a research prototype for compassionate debt recovery: customer and payment data are prepared, three AI branches produce distress and behavioral signals, a classifier assigns recovery tiers, and an audit rule can flag strategic defaults. Relationship managers explore the resulting customer-level decisions in a Streamlit dashboard. The README documents pipeline stages and capabilities whose runtime implementation is not present in the sampled source; those documented links are cited to the README, while dashboard data and manifest reads are evidenced in dashboard.py.
Main entry points: dashboard.py Streamlit dashboard · 04_feature_engineering.ipynb Feature engineering · 08_xgboost_recovery_tier.ipynb XGBoost classifier · 05_synthetic_chats.ipynb Synthetic chats
Components 15
Data preparation
- Source datasets
- Clean and merge
notebooks/Phase 1/03_dataset_merge.ipynb - Feature engineering
notebooks/Phase 1/04_feature_engineering.ipynb - Synthetic chats
notebooks/Phase 1/05_synthetic_chats.ipynb
AI signals
- FinBERT distress
notebooks/Phase 2/NLP/07_finbert_distress_score.ipynb - BiLSTM anomalies
notebooks/Phase 2/DL/06_bilstm_stress_patterns.ipynb
Recovery decisions
- XGBoost classifier
notebooks/Phase 2/ML/08_xgboost_recovery_tier.ipynb - Strategic-default audit
notebooks/Phase 2/ML/09_audit_layer.ipynb - Audited tier data
data/processed - Model manifest
models/xgboost/xgboost_manifest.json
Manager dashboard
- Streamlit dashboard
dashboard.py - Customer exploration
dashboard.py - Model explanations
dashboard.py
Other components
- Relationship manager
- OpenAI GPT-4o
Connections 16
- Source datasets → Clean and merge: clean and merge
README.md - Clean and merge → Feature engineering: engineer features
README.md - Feature engineering → Synthetic chats: generate chats
README.md - OpenAI GPT-4o → Synthetic chats: generate text
README.md - Synthetic chats → FinBERT distress: score distress
README.md - Feature engineering → BiLSTM anomalies: analyze payments
README.md - FinBERT distress → XGBoost classifier: fuse signals
README.md - BiLSTM anomalies → XGBoost classifier: fuse signals
README.md - Feature engineering → XGBoost classifier: provide features
README.md - XGBoost classifier → Strategic-default audit: classify then audit
README.md - Strategic-default audit → Audited tier data: write tier results
README.md - Relationship manager → Streamlit dashboard: opens
README.md - Streamlit dashboard → Audited tier data: reads CSV
dashboard.py - Streamlit dashboard → Model manifest: reads manifest
dashboard.py - Streamlit dashboard → Customer exploration: renders
README.md - Streamlit dashboard → Model explanations: renders
README.md