🏦 Credit Risk API — Full MLOps Pipeline

Full ML Engineering cycle for bank credit scoring: Training → MLflow Tracking → FastAPI → SHAP Explainability → Drift Monitoring → Docker

🔗 Live API: https://credit-scoring-ml-api.onrender.com/predict

---

📊 Model Performance

Metric	Value
ROC-AUC	Logged in MLflow per experiment
Algorithm	LightGBM + Class Imbalance handling
Decision threshold	0.15 (configurable)
Dataset	Home Credit Default Risk (Kaggle)
Explainability	SHAP TreeExplainer (Top-5 factors)

---

🛠 Tech Stack

Layer	Technology
Machine Learning	Python, Pandas, Scikit-Learn, LightGBM
Experiment Tracking	MLflow (autolog, Model Registry, SQLite backend)
Explainable AI	SHAP (TreeExplainer, Top-5 risk factors)
Backend	FastAPI, Uvicorn, Pydantic
Drift Monitoring	Evidently AI (Data Drift Detection)
Frontend	Streamlit (interactive scoring dashboard)
DevOps	Docker, Git
Audit Logging	SQLite (full prediction history)
Deployment	Render

---

⚙️ Architecture

📂 Home Credit Dataset (Kaggle)
        │
        ▼
🔬 src/train.py ──► MLflow autolog() ──► mlflow.db (SQLite)
        │                                      │
        │                               (metrics, ROC-AUC,
        │                                hyperparams, artifacts)
        ▼
🚀 api.py (FastAPI)
        │
        ├── lifespan: mlflow.lightgbm.load_model(RUN_ID)
        │            ← dynamic model loading from Registry
        │
        ├── /predict ──► SHAP TreeExplainer
        │            └──► Top-5 decision factors
        │            └──► log_request() ──► SQLite audit log
        │
        └── app.py (Streamlit UI) ──► visual scoring dashboard

📊 src/monitor_drift.py ──► Evidently AI ──► reports/data_drift_report.html

---

🔑 Key Features

1. MLflow Model Registry

Модель не "зашита" в код — при старте сервер динамически загружает нужную версию из MLflow по RUN_ID. Это позволяет переключаться между версиями модели без изменения кода API:

mlflow.set_tracking_uri("sqlite:///mlflow.db")
ml_models["lgbm"] = mlflow.lightgbm.load_model(f"runs:/{RUN_ID}/model")

2. Explainable AI (SHAP)

Каждое решение по кредиту сопровождается объяснением — топ-5 факторов которые повлияли на результат. Это требование банковских регуляторов (BASEL III):

{
  "probability_of_default": 0.73,
  "decision": "Reject",
  "explanation": [
    {"feature": "AMT_CREDIT", "impact": +0.42},
    {"feature": "DAYS_EMPLOYED", "impact": +0.31},
    {"feature": "AMT_INCOME_TOTAL", "impact": -0.18},
    {"feature": "DAYS_BIRTH", "impact": +0.15},
    {"feature": "EXT_SOURCE_2", "impact": -0.12}
  ]
}

3. Data Drift Monitoring (Evidently AI)

src/monitor_drift.py сравнивает референсные и текущие данные по 5 ключевым фичам модели и генерирует HTML-дашборд с алертами о дрейфе. Симулируется сценарий кризиса (рост доходов и кредитов в 3x):

python src/monitor_drift.py
# → reports/data_drift_report.html

4. Prediction Audit Log

Каждый /predict запрос логируется в SQLite: фичи клиента, вероятность дефолта, решение, timestamp. Полная воспроизводимость и аудируемость — обязательное требование для финансовых сервисов.

---

🚀 Quick Start

1. Prepare data

Download Home Credit Default Risk from Kaggle. Place application_train.csv and application_test.csv in data/raw/.

2. Install dependencies

pip install -r requirements.txt

3. Train model + MLflow tracking

python src/train.py

MLflow automatically saves hyperparameters, ROC-AUC metrics and model artifacts to mlflow.db.

4. View experiments in MLflow UI

mlflow server --host 127.0.0.1 --port 5000 --backend-store-uri sqlite:///mlflow.db

Open http://localhost:5000, copy the best RUN_ID and paste it into api.py.

5. Run API

# Local
uvicorn api:app --reload

# Docker
docker build -t credit-risk-api .
docker run -p 8000:8000 credit-risk-api

6. Run Streamlit UI

streamlit run app.py

7. Monitor Data Drift

python src/monitor_drift.py
# Open reports/data_drift_report.html in browser

---

📁 Project Structure

credit-risk-api/
├── src/
│   ├── train.py              # Training + MLflow autolog
│   ├── database.py           # SQLite prediction audit log
│   └── monitor_drift.py      # Evidently AI Data Drift
├── models/                   # Saved artifacts
├── notebooks/                # EDA and experiments
├── reports/                  # Evidently HTML reports
├── api.py                    # FastAPI microservice
├── app.py                    # Streamlit UI
├── Dockerfile
├── requirements.txt
└── README.md

---

🔗 Resources

• Home Credit Dataset on Kaggle
• MLflow Documentation
• Evidently AI Documentation
• SHAP Documentation

---

🔗 Related Projects

Part of a Fintech ML ecosystem:

• fraud-detection-api — Real-time fraud detection with Redis + A/B Testing
• fraud-gnn — Graph Neural Networks for fraud detection
• pfm-ai-assistant — Personal Finance Manager with AI advisor

💡 MLOps progression: this project covers the full cycle — training → experiment tracking → model registry → API → monitoring. That's exactly what senior ML Engineers do in production fintech systems.

---

📫 Author

Rashid Nurbekov — ML Engineer | Fintech & Generative AI | Almaty, Kazakhstan 🇰🇿