Inspiration
Financial crime costs the global economy trillions of dollars annually. In most countries, the process of reporting suspicious transactions from banks to regulatory authorities is still manual, slow, and error-prone — analysts fill out forms, attach documents, and email reports, sometimes days after a suspicious event occurs. I was inspired by a simple question: what if the moment a bank's system detects something suspicious, the right authority already knows about it? That real-time bridge between detection and action is what I set out to build.
What it does
Its a two-system financial intelligence platform: System 1 — FraudGuard (Institution Side) Banks and financial institutions use FraudGuard to analyze transactions in real time. The system uses machine learning (Isolation Forest + Random Forest) combined with rule-based detection to flag suspicious activity. A human analyst reviews flagged cases before submission. System 2 — Financial Intelligence Processing System (FRC Side) When a suspicious case is approved, it is automatically submitted via API to the central FRC platform. Here, the case is:
Assigned a unique FRC case ID Classified using structured legal rules (aligned with anti-money laundering legislation) Routed to the appropriate regulatory authority (e.g. financial intelligence units, tax authorities, law enforcement) Tracked through a full case lifecycle with reports, referrals, and audit logs
The result is a closed loop: detect → review → report → investigate → refer — all in one connected platform.
How we built it
Institution-side (FraudGuard)
Python + Flask backend with scikit-learn ML models Rule-based detection engine for threshold triggers Human review queue with AI-assisted narrative generation Deployed on Render
FRC Intelligence Platform
FastAPI backend with MongoDB JWT-based authentication with role-based access (admin, analyst, investigator, auditor) Institution registry with API key authentication Policy-driven intake API — submissions trigger automatically when regulatory thresholds are met, not by manual button clicks Structured legal knowledge base replacing raw PDF documents at runtime Report generation engine producing legally-grounded structured outputs Referral tracking to route cases to relevant authorities Full audit trail on every action Frontend built for the FRC dashboard, deployed on Vercel
Integration The two systems communicate through a secure REST API. FraudGuard submits a structured case payload to the FRC intake endpoint with full metadata, evidence references, and ML explanation. The FRC system acknowledges instantly and processes asynchronously.
Challenges we ran into
Designing the policy engine — deciding when submission should be automatic versus human-triggered required understanding how real financial regulations work and translating that into configurable backend logic Separating two systems cleanly — keeping FraudGuard and the FRC platform independent while making them communicate reliably without tight coupling was an architectural challenge Legal alignment without runtime PDF reading — extracting structured legal intelligence from regulatory documents and storing it as queryable rules rather than raw files took significant design thinking ML performance at scale — the original ML service struggled with large datasets; we implemented batch processing and summary-only responses to prevent timeouts and crashes Role-based access across two platforms — designing authentication that works for internal FRC users (human dashboard users) and external institution API clients required careful separation
Accomplishments that we're proud of
Built a fully functional two-system architecture that mirrors how real financial intelligence units operate Designed a policy-driven intake API — reporting is triggered by regulatory thresholds automatically, not by manual human action Created a structured legal knowledge base that powers case classification and report generation without depending on raw document parsing Delivered a complete case lifecycle: intake → review → legal classification → report → referral → audit trail Built an interface that genuinely looks and feels like a government-grade intelligence platform, not a student project
What we learned
Real-world regulatory systems are not just detection tools — the intelligence, legal classification, and routing layers are just as important as the ML models Designing for two separate user types (internal analysts and external institution API clients) requires fundamentally different thinking about authentication and access Structured data beats raw documents every time — building a legal rule engine from extracted Act summaries made the system far more powerful and maintainable than PDF-based approaches Scope discipline matters: a clean working demo of one complete flow is more impressive than ten incomplete features
What's next for AI POWERED SYSTEM FOR DETECTING SUSPICIOUS TRANSACTION
Multi-institution rollout — onboard multiple banks and financial institution types (SACCOs, payment providers, insurers) using the institution registry already built Advanced investigation ML — add cross-institution pattern detection, network analysis for linked accounts, and case prioritization scoring on the FRC side Full document generation — upgrade the report module to produce formally structured, legally formatted output documents ready for regulatory submission External agency integration — build authenticated APIs so downstream authorities (law enforcement, tax agencies, central banks) can receive and action referral packages directly Real-time monitoring dashboard — add live transaction stream visualization and live alert feeds for FRC analysts Expanded legal knowledge base — ingest the full set of anti-money laundering, tax, and financial crime legislation to support richer automatic case classification
Built With
- css
- fastapi
- flask
- html
- java
- javascript
- json
- jwt
- mongodb
- numpy
- pandas
- python
- react
- render
- rest
- scikit-learn
- tokens)
- typescript
- vercel
- web
Log in or sign up for Devpost to join the conversation.