Smart Banking platform

Banks constantly deal with the same four problems: who to lend to, which transactions to trust, where a currency is heading, and who their customers are — usually handled manually or with disconnected legacy tools instead of one system.

So we built a full-stack banking system where every core decision is backed by a trained model, not a static rule.

Backend — a Flask API with a model registry loading all four pipelines at startup, role-based auth, and one endpoint per model.

Frontend — two portals on the same backend: a User Portal for customers and a Manager Portal for staff, each tailored to that role.

Database — SQLite for true persistence instead of mocked data, securely storing accounts, logins, and every transaction and loan request, so the models evaluate real financial histories.

And at the center of it, four models, each doing real work:

Loan Approval — a customer applies, and an XGBoost classifier (best of 5 benchmarked algorithms — 93.1% accuracy, 0.978 ROC-AUC) returns an instant, SHAP-explained decision right on the manager's screen.

Fraud Detection — fraud is under 0.6% of transactions, so we evaluated on PR-AUC (0.96) instead of accuracy and tuned the threshold deliberately. A flagged transaction never moves money silently — it lands in the Manager's review queue with the probability and reasoning attached.

Currency Forecasting — a regression model trained on lag and moving-average features predicts the next trading day's EUR/USD close, giving both clients and managers a simple, data-driven read on where the currency is headed next.

Customer Segmentation — the one unsupervised piece: KMeans clustering groups customers into meaningful segments (Everyday Banking, Premium Relationship), giving managers a clear view of who they're banking.

The hardest part was never training a model in isolation — it's making all four behave consistently behind one backend and one UI, and actually deliver value to the user.

Proud to close this out with something that isn't just a model — it's a product.

A huge thank you to my incredible teammates for their hard work, support, and collaboration throughout this project: Ziad El Shamy and Zahraa Reda

Special appreciation to our instructors, Eng. Abdulrahman Bakeer and Eng. Mai Yasser, for their guidance, feedback, and continuous support — and for pushing us to build a product, not just a pipeline.

GitHub Repository: https://lnkd.in/eh-ZY6VU