FraudShield

FraudShield is a real-time fraud detection project for behavioral risk monitoring in online banking.

Comment

🛡️ FraudShield

Real-time fraud detection for online banking, powered by behavioral telemetry.

FraudShield combines classic transaction checks with session-level behavior signals to catch suspicious transfers earlier and give analysts clearer triage context.

Built with railtracks.

💡 Project Inspiration

Traditional fraud checks are often transaction-only (amount limits, destination checks, velocity). That leaves a gap: account takeover and automated abuse can look "normal" transaction-wise but still behave abnormally at the session level.

This project was motivated by the idea that fraud can be detected earlier by combining:

  • transaction context (amount, destination novelty, velocity), and
  • interaction behavior (navigation patterns, correction bursts, hesitation timing, mouse movement signals).

Why it matters in generative AI:

  • AI can help interpret noisy, multi-signal behavior patterns and improve analyst decision support.
  • Human analysts still stay in the loop with explainable factors, alerts, and case workflows.

🧱 Technology Stack

Languages

  • TypeScript
  • JavaScript
  • Python
  • CSS

Frameworks and Libraries

  • Next.js (App Router)
  • React
  • Tailwind CSS
  • MongoDB Node.js Driver
  • Playwright (concurrent behavioral harness)
  • Zustand (bank app state)
  • Railtracks (validation orchestration)

Platforms

  • MongoDB
  • HuggingFace OpenAI-compatible inference endpoint (optional AI co-assessment with gpt-oss-120b)
  • Localhost deployment for both apps (3000 and 3001)

Tools

  • Node.js / npm
  • Python / pip
  • Git / GitHub

🚀 Product Summary

FraudShield simulates a full fraud detection loop:

  1. A user performs banking actions in the banker app.
  2. Behavioral telemetry is captured and persisted.
  3. fraudshield scores sessions using rules and optional AI co-assessment.
  4. Analysts review live sessions, alerts, and cases.

User experience:

  • Customers interact with a realistic, synthetic banking flow in banker (accounts, transfer, activity).
  • Analysts use fraudshield to monitor sessions, inspect reason codes, triage alerts, and escalate cases quickly.
  • The dashboard emphasizes explainability so decisions are not based on opaque scores alone.

Core features:

  • Real-time risk scoring by session
  • Alert severity tiers (Low, Medium, High) with reason codes
  • Case workflow (Open, Investigating, Resolved)
  • Session drill-down with event timeline and top risk factors
  • Concurrent validation harness (up to 50 agents)
  • Run-scoped model comparison (rules-only vs rules+AI) for precision/recall/F1 and false-positive deltas

Innovative angle:

  • It blends transaction risk + behavior signals in one pipeline.
  • It supports explainable analyst workflows instead of black-box-only scoring.

🧭 Two Apps, One Flow

App Port Role
banker 3000 Synthetic banking client that generates transfer sessions + telemetry
fraudshield 3001 Analyst dashboard for scoring, alerts, and case workflow

📂 Repository Structure

banker/
  app/
  components/
  lib/
  testing/session-harness/
    simulate-session-batch.js
    README.md

fraudshield/
  app/
  components/
  lib/fraud/

testing/railtracks/
  run_ab_validation.py

⚙️ Setup

Prerequisites

  • Node.js 20+
  • npm
  • Python 3.10+
  • MongoDB reachable by both apps

Install

cd banker
npm install

cd ../fraudshield
npm install

cd ..
pip install -r requirements.txt

Environment

Create:

  • banker/.env (or .env.local)
  • fraudshield/.env

Minimum shared settings:

MONGODB_URI=mongodb://localhost:27017/
MONGODB_DB=northmaple_bank_demo

Optional FraudShield AI settings:

FRAUD_AI_ENABLED=true
FRAUD_AI_BASE_URL=haha nice try find ur own
FRAUD_AI_MODEL=openai/gpt-oss-120b
FRAUD_AI_API_KEY=haha nice try find ur own

▶️ Run Locally

Terminal 1:

cd banker
npm run dev

Terminal 2:

cd fraudshield
npm run dev

Open:

  • Bank UI: http://localhost:3000
  • Fraud Dashboard: http://localhost:3001/dashboard

🧪 Validation and Testing

50-Agent Concurrent Validation

From banker/:

node testing/session-harness/simulate-session-batch.js --phase=scale50 --capture=sample

Artifacts generated at runtime:

  • banker/testing/session-harness/latest/report.json
  • banker/testing/session-harness/latest/*.webm
  • banker/testing/session-harness/latest/*-activity.png

Verify Uplift for a Specific Run

Use:

http://localhost:3001/api/fraud/metrics?testRunId=<runId-from-report>

Key fields:

  • comparison.uplift.f1Delta
  • comparison.uplift.precisionDelta
  • comparison.uplift.falsePositiveRateDelta
  • comparison.aiAssessedSessions

Railtracks Runner

From repo root:

python testing/railtracks/run_ab_validation.py --phase=scale50 --total=50 --concurrency=10

Output:

  • testing/railtracks/latest-ab-summary.json

🤖 AI Use

Survey answer: Yes. More than 70% of implementation and iteration work was AI-assisted.

Built With

Share this project:

Updates