I have a sizeable collection of credit-card transaction logs currently stored in raw, unstructured form (JSON payloads, free-text notes, device metadata, etc.). My goal is to transform this data into a production-ready, supervised machine-learning model that reliably flags fraudulent activity. Here is what I need from you: • Parse and cleanse the unstructured records, engineer meaningful features, and align them with my existing “fraud / not-fraud” labels. • Train and fine-tune a supervised model—feel free to choose between scikit-learn, XGBoost, LightGBM, TensorFlow, or PyTorch as long as the final solution balances precision and recall and explains feature importance. • Produce an evaluation report that includes ROC-AUC, PR-AUC, confusion matrix, and threshold analysis so I can clearly see the trade-offs. • Package the model and preprocessing pipeline so I can load them in a single Python script or REST endpoint for real-time scoring. • Hand over clean, commented code plus a short README covering environment setup and retraining steps. I will consider the project complete once I can reproduce your metrics locally and run an inference call on a sample transaction.