DepositGuard — Bank Account Fraud Detection System

Back to Projects Featured Project

DepositGuard — Bank Account Fraud Detection System

Live Demo View code

1M+ Accounts Scored 0.89 AUC–ROC (XGBoost) 1.1% Fraud Prevalence 3-Stage Automated Pipeline Multi-Agent RAG Investigation

Architecture

Problem

Deposit account fraud, new accounts, ACH transfers, mule accounts, check fraud, costs U.S. banks billions every year.
Most banks still lean on legacy, rules-based detection: static logic that catches known patterns but misses new ones.
Manual investigation against dense regulation (Reg E, NACHA, BSA/AML) is slow, inconsistent across analysts, and hard to audit at scale.

Approach

An XGBoost model, benchmarked against three other classifiers under realistic 1:90 fraud imbalance, scores every account and explains its call with SHAP.
High-risk accounts are handed to AI agents that investigate against real regulatory text (Reg E, NACHA, BSA/AML), not general knowledge.
A fixed, inspectable rule set, not the AI, decides the final outcome, keeping every escalation auditable.

Impact

A complete, live fraud detection platform, not a static notebook or one-off demo.
Every account gets an explainable risk score, a grounded investigation trail, and a documented decision, all inspectable end to end.
Ships as a live analytics dashboard, so model performance and case investigations stay explorable in real time, not locked in a report.

How It Works

1. Risk Scoring

Every account gets a fraud-risk score from an XGBoost model trained on real behavioral signals, like device type, email type, and how fast the application moved. Each score comes with its top contributing factors, not just a number.

2. Investigation

High-risk accounts are handed to AI agents that investigate against actual banking regulations, Reg E, NACHA, BSA/AML, pulled from a live knowledge base rather than general knowledge, so the reasoning stays grounded in real policy.

3. Decision Gate

A fixed rule set, not the AI, makes the final call. Freeze, escalate, monitor, or clear, every decision traces back to a documented, inspectable rule, not a model's opinion.

Results at a Glance

XGBoost led all four models tested, scoring 0.89 AUC–ROC on unseen data it never trained on.
Device type and email type are the strongest individual red flags the model relies on.
0.5% of accounts are flagged high risk and routed for investigation; 96.5% clear automatically.

Tech Stack

Python XGBoost LangGraph SHAP ChromaDB (RAG) Gemini & Groq Streamlit