Real-Time Fraud Detection Model

I am working with the PaySim transaction dataset from Kaggle and need an extra pair of expert hands to transform it into a robust, real-time fraud-detection pipeline. The core objective is clear: spot fraudulent transfers the moment they appear, not minutes or hours later.

Where the project stands
– I have the raw PaySim CSVs plus my own exploratory notes.
– Basic cleaning is done; now the heavy lifting—feature engineering for real-time scoring—needs to happen.
– I also want a repeatable test harness so we can prove the model’s speed and accuracy before deployment.

What I need from you
1. Thoughtful feature engineering that balances predictive power with low-latency inference.
2. A supervised model (tree-based ensemble, deep network, or anything you justify) tuned for real-time fraud detection.
3. End-to-end Python code (pandas, scikit-learn or TensorFlow/PyTorch) that:
• ingests live-like streams or mini-batches,
• applies your engineered features,
• outputs a fraud score instantly, and
• logs the decision for auditing.
4. A concise test suite—preferably in pytest or a Jupyter notebook—that simulates incoming transactions and reports precision, recall, and latency per batch.

Acceptance criteria
– <250 ms median prediction time on a standard laptop CPU.
– ROC-AUC >= 0.95 on the official PaySim hold-out split.
– Code documented and ready to drop into a REST or Kafka microservice later.

If this matches your skill set, let’s discuss your proposed feature ideas, model choice, and how quickly you can spin up an initial proof-of-concept.