The idea for FraudShield AI came from experiencing the two extremes of modern banking failures: the frustration of an innocent card decline while traveling or buying coffee, contrasted with the absolute nightmare of dealing with actual payment fraud. When unauthorized charges happen, victims face 30 to 90 days of call center escalations, branch visits, and manual paperwork. We asked ourselves: Why can banks process a million transactions a second, yet take 45 days to resolve an obvious dispute? We set out to build an embedded system that intercepts fraud in sub-seconds without locking out genuine cardholders, while closing dispute cycles autonomously.

  What It Does

FraudShield AI is an in-flight fraud interception and autonomous dispute resolution engine:

Zero False-Block Interception: High-risk anomalies trigger an interactive, one-tap mobile push prompt ("Did you authorize \$3,450 at LuxuryElectronics_Dubai?") instead of an embarrassing hard freeze.
Instant Legal Dispute Dossier: If the user flags fraud, an autonomous agent pipeline immediately revokes the payment token, auto-maps network scheme rules (Visa Reason Code 10.4: Card-Absent Fraud), and compiles a legally compliant bank dispute packet with full telemetry evidence in under 30 seconds.

How We Built It
Frontend / PWA: Developed a mobile-first interface using Next.js 14, Tailwind CSS, and Lucide Icons featuring a live alert banner and a multi-step resolution stepper.
Real-Time Backend: Engineered a high-throughput FastAPI service communicating with the frontend over WebSockets for sub-second bidirectional telemetry.
Analytics Engine: Utilized DuckDB for in-memory columnar evaluation of geo-velocity deltas (impossible travel hops) and merchant risk scoring.
Agentic AI & PDF Engine: Integrated the Gemini API to analyze transaction switch logs, map network reason codes, and compile formal bank-ready PDF dispute packets using ReportLab.
```
Challenges We Faced
```
Sub-Second Latency: Ingesting high-velocity card swipe streams and calculating geo-distance deltas without blocking the real-time WebSocket connection. We solved this with DuckDB's in-memory engine and asynchronous execution.
State Synchronization: Managing the real-time lifecycle between transaction interception, mobile push verification, and token revocation without dropped events.
Scheme Compliance: Structuring automated dispute dossiers to match strict payment scheme standards (Visa/Mastercard reason codes and ISO 8583 logs) so the output is immediately actionable by card-issuing banks.

What We Learned We gained deep insights into core banking rails, chargeback regulations, and real-time event streaming. Most importantly, we learned that the future of FinTech security isn't just about passive ML anomaly detection—it lies in pairing real-time customer-in-the-loop verification with autonomous agentic execution to eliminate weeks of bureaucratic paperwork.