Causal AI credit risk assessment for East African digital lending markets.
Built for thin-file borrowers, M-Pesa ecosystems, and CBK regulatory compliance.
CreditPulse is an open-source credit risk intelligence platform that goes beyond correlation-based scoring. It asks not just who will default, but why — using causal inference, Bayesian uncertainty quantification, and explainable AI.
| Model | Metric | Value |
|---|---|---|
| XGBoost Classifier | AUC-ROC | 0.7415 |
| XGBoost Classifier | Gini Coefficient | 0.4830 |
| Random Survival Forest | C-index | 0.8130 |
| Difference-in-Differences | Causal coefficient | -0.1125 (p=0.002) |
| Regression Discontinuity | LATE at cut-off | 0.0058 (p=0.0028) |
| Fraud Detector | Applications flagged | 1.75% |
# 1. Clone
git clone https://github.com/B-Omare/creditpulse.git
cd creditpulse
# 2. Environment
conda create -n creditpulse python=3.11 -y
conda activate creditpulse
pip install -e .
pip install pandas numpy pyarrow xgboost==2.1.4 scikit-learn joblib shap
pip install scikit-survival dowhy econml statsmodels networkx matplotlib seaborn
pip install langchain langchain-community langchain-text-splitters
pip install langchain-huggingface chromadb bertopic sentence-transformers
pip install fastapi uvicorn streamlit pydantic httpx
# 3. Run the pipeline
python creditpulse/ingestion/clean.py
python creditpulse/causal/dag.py
python creditpulse/causal/diff_in_diff.py
python creditpulse/causal/regression_discontinuity.py
python creditpulse/nlp/rag_pipeline.py
python creditpulse/nlp/topic_model.py
python creditpulse/models/xgboost_model.py
python creditpulse/models/bayesian_model.py
python creditpulse/models/survival_model.py
python creditpulse/models/fraud_detector.py
python creditpulse/explainability/fairness_audit.py
python creditpulse/explainability/model_card.py
# 4. Launch
# Terminal 1
uvicorn creditpulse.api.main:app --reload
# Terminal 2
streamlit run app/streamlit_app.pyThen open:
- Dashboard: http://localhost:8501
- API docs: http://127.0.0.1:8000/docs
For full reproduction instructions see REPRODUCE.md.
| Layer | Component | Technology |
|---|---|---|
| Data | Ingestion & ETL | pandas + Snakemake |
| Data | IFRS 9 Schema | Domain logic |
| Causal | Causal DAG | DoWhy + networkx |
| Causal | DiD Experiment | statsmodels |
| Causal | Regression Discontinuity | statsmodels + matplotlib |
| NLP | RAG Chatbot | LangChain + ChromaDB |
| NLP | Topic Modelling | BERTopic |
| NLP | Decision Explainer | Rule-based engine |
| ML | XGBoost Classifier | XGBoost + SHAP |
| ML | Bayesian Scorer | Bootstrap ensemble |
| ML | Survival Model | scikit-survival |
| ML | Fraud Detector | Isolation Forest |
| Explainability | Fairness Audit | sklearn metrics |
| Explainability | Model Card | SR 11-7 Markdown |
| API | Scoring Service | FastAPI + uvicorn |
| UI | Dashboard | Streamlit |
Score individual borrowers with interactive sliders. Returns default probability, IFRS 9 stage, Bayesian credible interval, fraud flag, and a plain-language explanation compliant with CBK 2022 regulations.
Portfolio-level IFRS 9 stage distribution, key risk metrics, and a causal scenario modelling tool — estimate the portfolio impact of a macro shock using the DiD coefficient.
Fairness audit across borrower segments, causal evidence summary (DiD + RD), and the full SR 11-7 model card.
| Standard | Status |
|---|---|
| IFRS 9 (loan stage classification) | Implemented |
| CBK Digital Credit Regulations 2022 | Borrower explanations implemented |
| SR 11-7 (model risk management) | Model card generated |
| Fairness audit (age, income, occupation) | Implemented |
@software{omare2026creditpulse,
author = {Omare, Brian},
title = {CreditPulse: Causal AI Credit Risk Assessment
for East African Digital Lending Markets},
year = {2026},
url = {https://github.com/B-Omare/creditpulse},
version = {0.1.0}
}docker-compose up --buildRequires 4GB+ free RAM. Starts both the API (port 8000) and dashboard (port 8501) in separate containers.
Built with Python 3.11 | Tested on Windows 11 and Ubuntu 24.04