End-to-end credit risk scoring application built with:
- XGBoost
- FastAPI
- Streamlit
- MLflow
- Docker
This project demonstrates the development of an end-to-end credit risk data product.
Starting from raw credit portfolio data, the solution performs feature engineering, trains and evaluates machine learning models, tracks experiments with MLflow, serves predictions through a FastAPI REST API and exposes a business-friendly Streamlit dashboard.
The objective is to estimate Probability of Default (PD) for credit applicants or existing portfolio customers while showcasing the full lifecycle of a modern machine learning solution.
The project focuses on practical data science and machine learning engineering concepts, including:
- Feature engineering
- Model experimentation
- Model selection
- Experiment tracking
- API development
- Dashboard development
- Containerization
- End-to-end deployment
Version 0.1 follows an experimentation-first approach, allowing feature groups and data sources to be evaluated incrementally before moving toward a production-grade architecture.
Raw Data
│
▼
Feature Engineering
│
▼
XGBoost Training
│
▼
MLflow Tracking
│
▼
FastAPI REST API
│
▼
Streamlit Dashboard
Version 0.1 was intentionally built as an experimentation-focused release.
Instead of building a fully optimized production pipeline from day one, feature blocks were added incrementally to understand their contribution to model performance and business value.
The workflow followed:
- Static application features
- Person-level aggregates
- Previous application features
- Bureau B aggregates
- Bureau A1 lightweight features
- Bureau A2 lightweight features
Each feature block was merged, validated, and evaluated independently before moving to the next stage.
This approach improved explainability of the development process and made it easier to identify which data sources contributed the most predictive power.
The largest performance improvements came from Bureau A1 and Bureau A2 delinquency-related features, particularly overdue amount and payment delay aggregates.
Algorithm:
- XGBoost Classifier
Dataset size:
- ~1.5 million observations
Feature count:
- 258 engineered features
Performance:
| Metric | Value |
|---|---|
| Validation AUC | 0.824 |
| Validation Gini | 0.648 |
- XGBoost
- Scikit-Learn
- Pandas
- NumPy
- FastAPI
- Swagger UI
- Uvicorn
- Streamlit
- MLflow
- Docker
Start FastAPI:
uvicorn api.main:app --reloadSwagger Documentation:
http://localhost:8000/docs
Example API Workflow:
Applicant Data
│
▼
POST /predict
│
▼
Probability of Default
│
▼
Risk Band
Run locally:
streamlit run app_streamlit/app.pyThe dashboard communicates with the FastAPI backend and provides a simple business-facing interface for risk scoring.
Launch MLflow UI:
mlflow uiOpen:
http://localhost:5000
MLflow is used to track:
- Validation metrics
- Model versions
- Feature counts
- Experiment metadata
Build and run:
docker compose up --buildLarge local artifacts are intentionally excluded from Git:
- Raw datasets
- Processed parquet files
- Trained model binaries
- MLflow databases
- MLflow run artifacts
To reproduce the project, execute the scripts inside the scripts/ directory.
Version 0.1 focuses on:
- Local experimentation
- Feature engineering
- Model comparison
- API deployment
- Dashboard deployment
- MLflow integration
The primary goal was to understand the end-to-end workflow before investing in production-grade optimization.
Planned improvements:
- Hyperparameter optimization
- SHAP explainability
- Model monitoring
- Drift detection
- Automated retraining
- CI/CD pipeline
- Cloud deployment
- Production-grade feature pipelines
- Enhanced Streamlit user experience