This project takes a fraud-detection

This project takes a fraud-detection model built in a notebook and turns it into a small, runnable system you can demo and deploy locally. In my workflow, I start with the notebook, save a single artifact, then wrap it with an API, a UI, and basic monitoring.

1. Train and evaluate a model on creditcard.csv

2. Save a deployable artifact (.pkl)

3. Serve predictions through a FastAPI backend

4. Provide a Streamlit UI for manual testing and batch scoring

5. Log predictions and feedback labels for monitoring

6. Run basic drift checks and export feature importance for sanity checks

What Problem This Solves

Fraud detection is a highly imbalanced classification problem, so "accuracy" is usually misleading. What you typically want is:

1. a risk score (fraud probability) per transaction

2. a decision policy (threshold) you can tune to match operational goals (precision/recall tradeoff)

3.  serving layer so the model works outside Jupyter

4. a monitoring loop so you can detect drift and decide when to retrain