Important

Project Status: This project is currently a Work in Progress.

Warden: Distributed Real-Time Fraud Detection System

Warden is a high-performance fraud detection ecosystem designed to identify anomalous transaction patterns through distributed orchestration and machine learning. It covers the full lifecycle from synthetic data generation to explainable ML-driven inference.

Architecture Overview

The system is decoupled into three specialized services to ensure scalability and fault tolerance:

Warden UI (React): Real-time monitoring dashboard featuring synthetic data generation pipelines, pre-simulation feature analytics, and live model insight logs (with post-simulation analytics).

Warden Backend (Spring Boot): The central orchestration layer. Manages PostgreSQL/Redis state, executes complex feature engineering, ensures transaction idempotency, and maintains MQ reliability.

Warden ML (FastAPI): A Python-based intelligence service utilizing Gradient Boosting models to provide probabilistic fraud scores and basic human-readable reasoning.

Key Features

1. Data Generation & Pipelines

Adversarial Simulation: Custom pipelines to generate realistic "Attack" scenarios, including Impossible Travel (high-speed GPS shifts), Behavioral Anomalies, Temporal Deviations (odd-hour transactions mapped to local timezones), and Statistical Surges.

Feature Analytics: Integrated pre-simulation tools to analyze feature distributions before they hit the inference engine.

2. High-Performance Orchestration

Feature Engineering: Real-time calculation of ATV Delta, Geodesic Distance, and Temporal Velocity using Redis for low-latency state lookups.

Reliability & Idempotency: Custom logic to prevent duplicate processing and a 30-second watchdog timer that re-queues lost transactions to ensure 100% processing uptime.

3. Explainable ML Inference

Probabilistic Scoring: Moves beyond binary classification to provide a raw fraud probability (0.0 to 1.0).

Model Explainability: Generates basic natural language reasons for every decision.

Real-time Loop: Inferences are pushed back to the UI via the backend for immediate visualization, logging, and post-simulation analytics.

Tech Stack

Backend: Spring Boot, Java, PostgreSQL, Redis

Machine Learning: Python (FastAPI), Scikit-Learn / XGBoost

Message Broker: RabbitMQ

Frontend: React.js, Recharts, and Nivo (for visualization of feature analytics)