Credit Risk Engine

An end-to-end, containerized microservice for real-time credit default prediction.

Built on the 300,000+ row Home Credit Default Risk dataset, with Bayesian hyperparameter tuning (Optuna), MLflow experiment tracking, and a containerized FastAPI inference endpoint.

Project Origin: The idea for this project emerged while completing IBM's AI Fundamentals course, where I explored IBM Watson's automated machine learning workflow. That experience motivated me to build a similar credit risk pipeline from scratch, implementing the data engineering, model training, optimization, deployment, and monitoring components myself.

System Architecture

The pipeline is entirely modular and config-driven, built with the following stack:

• Data Engineering: Polars for lazy-evaluated, memory-efficient relational joins across multiple banking tables.
• Machine Learning: XGBoost utilizing native GPU acceleration (device: cuda) and class-weighting (scale_pos_weight) to mathematically penalize false negatives.
• Bayesian Optimization: Optuna replaces brute-force grid search, utilizing probabilistic models to navigate the hyperparameter space efficiently.
• MLOps Tracking: MLflow integrated via Optuna callbacks to log trials, artifacts, and evaluation metrics to a local SQLite backend.
• Explainability: SHAP (SHapley Additive exPlanations) for compliance-ready, tree-based feature importance mapping.
• Automated Risk Rationale: LLM integration (Llama-3.1) to dynamically generate clinical, natural-language rationale for credit rejections.
• Deployment: FastAPI and Docker to wrap the predictive model in a highly available REST API.
• CI/CD & DevOps: GitHub Actions workflow for automated Docker builds and image publishing to the GitHub Container Registry (GHCR).

Training Results & Evaluation

Hyperparameter optimization was executed using Optuna Bayesian Optimization over the XGBoost space.

Test Set Evaluation (30,752 samples):

• Overall Accuracy: 86.4%
• Default Recall: 38.9% (Successfully flagged ~39% of all actual defaults in a highly imbalanced environment where only ~8% of total applications default)
• Default Precision: 26.6%
• F1-Score (Default Class): 0.316

MLOps Dashboard

To view the tracking dashboard locally:

mlflow ui --backend-store-uri sqlite:///mlruns.db

Local Setup & Deployment

Prerequisites: Create a .env file in the root directory and add your Groq API key:

GROQ_API_KEY=your_api_key_here

Install dependencies:

pip install -r requirements.txt

Fetch data and run the training pipeline:

python src/download_data.py
python main.py

Serve the Model (FastAPI):

uvicorn src.api:app --reload

Navigate to http://127.0.0.1:8000/docs to test the API via the interactive Swagger UI, or use the provided utility script to run a local inference test:

python scripts/test_api_inference.py

Run via Docker:

docker build -t credit-risk-engine .
docker run -p 8000:8000 credit-risk-engine