ranji-t/fraud-detection-ml-system — reverse-engineered prompt

Build me a portfolio ready credit card fraud detection project that shows the whole machine learning workflow, from loading the Kaggle credit card fraud data to training models, evaluating them, explaining predictions, and serving results in a reproducible way.

I want it to compare simple models with a neural network built from scratch in JAX, handle the very rare fraud cases properly, and focus on practical fraud review metrics like precision at K, recall at K, lift at K, ranking quality, precision recall curves, confusion matrices, and threshold tuning. Add an automated tuning step that can balance recall and precision instead of optimizing only one score.

Please include clear notebooks for exploration and experiments, reusable Python code for metrics and training, SHAP style explanations for why a transaction looks risky, config files, tests, formatting checks, Docker support, and a simple inference service or demo endpoint. Keep it clean enough that I can push it to GitHub and show it as an end to end ML engineering project.