Benderrrrr/credit-scoring

Credit Scoring - ML Application

Purpose

This project predicts credit default risk from Home Credit data and exposes the model through a small production-style stack: a FastAPI backend, a Streamlit application, monitoring views, and CI/CD deployment to Hugging Face Spaces.

The current app is focused on serving and monitoring a selected production model. Some training options are configurable in code and YAML, but they are not exposed as a fully dynamic product interface.

Data And Model Scope

The preprocessing pipeline combines the main application data with historical credit and previous application tables:

• —application train/test data;
• —bureau and bureau balance history;
• —previous applications, installments, POS cash and credit card balance data.

The production model currently served by the API is a LightGBM classifier using a reduced set of 20 selected features. The training code can instantiate several model families from configuration, including LightGBM, XGBoost, CatBoost, random forest, logistic regression and a dummy baseline.

Key Features

• —Training pipeline: YAML-driven experiment setup with preprocessing, train/test split, threshold optimisation and MLflow logging.
• —Feature selection: stability-oriented feature ranking using repeated LightGBM importance over folds and random seeds.
• —FastAPI serving: prediction, client lookup, model threshold and reference-data routes, with documented errors in Swagger.
• —Streamlit scoring UI: load a client, edit the selected model features, run predictions and compare each value with the reference distribution.
• —Streamlit monitoring UI: API route monitoring, latency/error summaries, Evidently data drift and data quality reports.
• —Inference optimisation: ONNX export and Streamlit benchmark comparing standard LightGBM inference with ONNX Runtime.
• —CI/CD: GitHub Actions run Ruff and Pytest, then deploy to Hugging Face Spaces after a successful CI run on main.

Current Limitations

• —The Streamlit app does not let users dynamically choose data sources or train new models from the UI.
• —The deployed app serves the packaged production model and reference parquet files.
• —The broader model registry exists in training code, but the current serving path is built around the selected LightGBM model.

Project Structure

src/credit_scoring/
|-- features/      # preprocessing and feature selection
|-- models/        # training, evaluation, tuning and explainability
|-- serving/       # FastAPI app, inference code and packaged model assets
`-- interfaces/    # Streamlit prediction and monitoring pages

scripts/
|-- export_onnx.py
|-- generate_base_for_inference.py
|-- run_ft_selection_nb.py
`-- run_ft_selection_ranking.py

Main Commands

just api          # FastAPI on port 8000
just dashboard    # Streamlit app
just app          # API and dashboard together
just train        # training pipeline
just export-onnx  # export the production model to ONNX
just test         # test suite

Tech Stack

• —Backend: FastAPI, Pydantic, Uvicorn.
• —Frontend: Streamlit and Plotly.
• —ML: Pandas, Scikit-learn, LightGBM, ONNX Runtime, MLflow.
• —Monitoring: API JSONL logs, Evidently reports, Streamlit dashboards.
• —Automation: Ruff, Pytest, GitHub Actions, Docker and Hugging Face Spaces.