AI-powered financial anomaly Detection

An AI-powered financial anomaly detection system that reduces investigator false-positive fatigue by 4x using ensemble machine learning and plain-English LLM explanations.

Comment

FraudOS (AEGIS.AI)

Inspiration

Today, financial fraud detection pipelines flag transactions at an average 92% false-positive rate. This means that out of every 100 transactions an investigator manually reviews, 92 are legitimate customers being inconvenienced, and only 8 are actual fraud cases.

This creates a massively noisy queue, leading to:

  • Investigator fatigue
  • Missed fraudulent activity
  • Frustrated customers
  • Slower operational response times

We were inspired to solve this problem of “alert fatigue” by building a system that not only detects fraud more accurately, but also explains why the AI flagged a transaction — eliminating the traditional “black box” problem in financial machine learning systems.

What it does

FraudOS (AEGIS.AI) is an end-to-end fraud investigation operating system powered by our virtual AI agent, “Rahul.”

🔍 Hybrid AI Detection

The platform ingests live transaction data and scores activity using an Ensemble AI Model that combines:

  • Isolation Forest → detects point anomalies
  • LSTM Networks → detects sequential and velocity-based fraud patterns

This hybrid approach improves fraud detection accuracy while reducing false positives.

🧠 Explainable AI (XAI)

When a transaction crosses the anomaly threshold (for example, > 0.80), Agent Rahul generates a plain-English explanation for investigators.

Example:

CRITICAL: High-value wire transfer of $34,200 initiated from a new device in Hong Kong.

This removes the black-box nature of AI and allows investigators to quickly understand the reasoning behind each alert.

👨‍💼 Human-in-the-loop Governance

Investigators interact with a premium desktop-like web interface where they can:

  • Review flagged transactions
  • Confirm fraud
  • Clear legitimate activity
  • Monitor compliance metrics

The platform also enforces regulatory compliance by automatically blocking any model threshold changes that would reduce fraud recall below 95%.

📡 Live Data Injection

FraudOS includes a real-time Python AI Data Agent that continuously simulates realistic transaction activity and injects it directly into the cloud database for live monitoring and testing.

How we built it

🎨 Frontend (UI/UX)

Built using React + Vite and deployed on Vercel.

We designed a highly immersive glassmorphic Web Operating System featuring:

  • macOS-style dock
  • Draggable application windows
  • Smooth micro-interactions
  • Real-time toast notifications
  • Interactive fraud investigation workspace

The goal was to make enterprise fraud analysis feel modern, intuitive, and visually premium.

⚙️ Backend (FastAPI)

A fully containerized Python FastAPI backend, deployed on Render, powers:

  • Business logic
  • Fraud scoring APIs
  • Compliance workflows
  • Real-time metrics
  • Automated PDF compliance report generation

🗄️ Database (PostgreSQL)

We migrated from local storage to a cloud-hosted Supabase PostgreSQL database to support:

  • Persistent live state
  • Instant audit logging
  • Scalable cloud infrastructure
  • Reliable multi-service synchronization

🤖 AI Data Agent

We developed a custom Python worker script (ai_data_agent.py) using:

  • scikit-learn
  • Faker

The agent continuously injects realistic synthetic transaction data into Supabase, enabling a constantly active fraud monitoring environment.

Challenges we ran into

One of the biggest technical hurdles was managing asynchronous communication between:

  • The live AI data injector
  • The cloud database
  • The frontend dashboard

Ensuring that the Live Activity Feed updated accurately whenever the Python worker injected new transactions required careful handling of:

  • Database connection pooling
  • asyncpg optimization
  • Supabase free-tier connection limitations

We also had to properly configure CORS across separate Vercel and Render deployments to ensure secure cross-origin communication.

Accomplishments that we're proud of

✨ Exceptional UI/UX

We didn’t build a traditional dashboard — we created a fully interactive Web Operating System with premium animations, glassmorphism, and enterprise-grade usability.

☁️ True Full-Stack Cloud Deployment

Successfully deploying and integrating:

  • Vercel
  • Render
  • Supabase
  • Containerized FastAPI services

was a major milestone for the team.

🛡️ Regulatory-First Design

FraudOS actively enforces compliance by:

  • Blocking unsafe model threshold modifications
  • Maintaining minimum fraud recall guarantees
  • Automatically generating compliance PDFs and audit logs

What we learned

We learned that even a highly accurate ML model is ineffective in finance if human investigators cannot understand why decisions are being made.

The most valuable insight was realizing that:

Explainability is just as important as accuracy in enterprise AI systems.

Bridging machine learning with human-centered UX through plain-English AI explanations became the defining strength of FraudOS.

What's next for FraudOS (AEGIS.AI)

⚡ Real-time WebSocket Integration

We plan to move from REST API polling to live WebSocket architecture so investigators receive updates instantly the moment suspicious activity is detected.

🧩 Custom Agent Personas

Future versions of Agent Rahul will support specialized personas tailored to different fraud domains, such as:

  • Crypto Fraud Specialist
  • Domestic Retail Fraud Analyst
  • International Wire Transfer Expert
  • Regional Compliance Specialist

This will allow financial institutions to customize the AI investigator experience based on operational needs.

Tech Stack

Frontend  : React, Vite, TailwindCSS
Backend   : FastAPI, Python
Database  : Supabase PostgreSQL
AI/ML     : Isolation Forest, LSTM, scikit-learn
Cloud     : Vercel, Render
Utilities : Faker, asyncpg

Final Vision

FraudOS (AEGIS.AI) is more than a fraud dashboard — it is an intelligent fraud investigation operating system designed to combine:

  • AI-powered detection
  • Human-centered explainability
  • Regulatory governance
  • Real-time operational workflows

Our mission is to dramatically reduce alert fatigue while empowering investigators with transparent, trustworthy AI.

Built With

Share this project:

Updates