FraudWatch (Python, FastAPI + scikit-learn)

Mini piattaforma MLOps: training di un modello di classificazione fraud/no-fraud, servizio di prediction via API, test, Docker e CI.

Disclaimer This repository is an educational demo for end-to-end ML (training + serving). The model is trained on synthetic data and must not be used to make real financial decisions. No real personal data is included.

Quick start (local)

python -m venv .venv && source .venv/bin/activate  # (Windows: .venv\Scripts\activate)
pip install -r requirements.txt
python training/train.py         # genera models/model.pkl
uvicorn app.main:app --reload    # http://localhost:8000/docs

Endpoints

• GET /health → stato
• POST /predict → body: { "features": [0.1, 0.2, 0.3, ...] }

Test

pytest -q

Docker

docker build -t fraudwatch-api .
docker run -p 8000:8000 fraudwatch-api
# oppure
docker compose up --build

Framework

fraudwatch-python/
  app/
    main.py
    routers/predict.py
    schemas.py
  training/train.py
  models/ (generata)
  tests/test_api.py
  requirements.txt
  Dockerfile
  docker-compose.yml
  .github/workflows/python-ci.yml
  Makefile

Limitations & Responsible Use

• Synthetic dataset; results are illustrative only.
• No calibration/monitoring/drift checks included yet.
• No PII is processed; do not submit real customer data to the demo API.
• Before any production use, add: data governance, monitoring, bias audits, rate limiting, auth, logging, model registry (e.g., MLflow).

---

© drfb02