{"repo":"Md-Emon-Hasan/FraudChurn-Nexus","free":true,"listed":false,"github":"https://github.com/Md-Emon-Hasan/FraudChurn-Nexus","clone":"git clone https://github.com/Md-Emon-Hasan/FraudChurn-Nexus.git","description":"Unified ML platform serving two production risk models behind one API, fraud detection using a Logistic Regression pipeline at 92.4 percent accuracy and 0.90 F1, and churn prediction using a five-estimator hard-voting ensemble at 85.6 percent accuracy, with SMOTE balancing, sub-0.5 second inference, and CI-enforced 90 percent test coverage.","language":"Jupyter Notebook","stars":10,"topics":["anomaly-detection","classification","data-science","ecommerce","financial-data","fraud-detection","fraudulent-transactions","machine-learning","online-shop","risk-management"],"license":null,"category":"machine-learning","readme_excerpt":"FraudChurn Nexus: Unified Fraud and Churn Prediction Online stores discover fraudulent orders weeks later, when the chargeback arrives; telecom operators find out a customer was unhappy only after the cancellation goes through. Both problems are usually handled the same way — an analyst reading through transaction records and account histories one row at a time, deciding by gut feel. FraudChurn Nexus replaces that manual screening. Enter an order or a customer account and it returns a clear verdict along with how confident it is , so borderline cases are visible instead of hidden, and every decision is written to a permanent record you can audit later . Review teams stop hand-checking routine cases and spend their time only where the judgement call is genuinely close. Under the hood it is a production-ready monolith : a FastAPI backend with Pydantic -validated schemas serving two independently trained scikit-learn engines — a five-estimator hard-voting ensemble (Gradient Boosting, AdaBoost, Random Forest, Decision Tree, Logistic Regression) for churn, trained on SMOTE -rebalanced data to correct severe class imbalance, and a one-hot-encoded Logistic Regression pipeline for fraud — both loaded from pickled artifacts at startup, with every request and result logged to SQLite . Form dropdowns are derived from the training data itself, so the UI cannot submit a category the model has never seen. A React 19 + Vite + Tailwind/DaisyUI glassmorphism frontend, rotating-file logging, a","default_branch":null,"files":null,"tree":[],"storefront":"/r/Md-Emon-Hasan","claimed":false,"request_supported":{"post":"https://gitbuyer.com/r/Md-Emon-Hasan/FraudChurn-Nexus/request-supported","requests":0},"note":"indexed from public GitHub; nothing is for sale on this page. Clone it from GitHub. Paid listings live at /search."}