Back to projects
Jan 01, 2026
2 min read

ScamIntelli

Real-time fraud detection platform powered by graph machine learning. Detects coordinated fraud rings by analyzing entity relationships, shared attributes, and transaction topology using GNNs and gradient boosting.

Overview

ScamIntelli is a real-time fraud detection system that goes beyond per-transaction analysis to detect coordinated fraud rings by reasoning about the relationships between entities — accounts, devices, merchants, and IP addresses.

Architecture

The system is composed of three layers:

Ingestion: A FastAPI service validates incoming transaction events using Pydantic, enriches them with velocity and reputation features, and publishes them to a Redis Stream.

Graph Engine: A worker pool consumes events and maintains a live property graph in PostgreSQL. Nodes represent entities; edges represent transactions, logins, and shared attributes. Hot subgraphs are cached in Redis for low-latency lookup.

Detection: For each transaction, the engine extracts a k-hop neighborhood subgraph, computes structural features (centrality, clustering, shared neighbors), and feeds them into a GNN + gradient boosting ensemble. Predictions include a calibrated confidence score and an attribution trace pointing to the specific graph features that drove the decision.

Key Results

  • ~23% improvement in fraud ring detection vs. tabular-only baseline
  • Sub-100ms prediction latency for most transactions (graph cache hit)
  • Full attribution trace for compliance and human review

Tech Stack

Python · FastAPI · PostgreSQL · Redis · PyTorch Geometric · Docker · Pydantic