Graph-Based Fraud Detection

A living fraud intelligence graph built to expose laundering rings, funnel accounts, and coordinated abuse invisible to flat models.

Open repository

Project claim

Expose laundering rings and coordinated transaction abuse.

Role

Graph workflow designer

Focus signals

graph featuresfraud signalsrelational reasoning

Proof signal 1

31.9M transactions modeled

Proof signal 2

Palantir-inspired ontology framing

Proof signal 3

Dashboard + LLM investigation surface

Challenge

Traditional fraud models miss relational behavior, especially when fraud hides across users, devices, and transaction graphs.

Solution

Designed a graph-based fraud workflow that turns transactions into connected behavioral signals, then uses graph-aware features and ML reasoning to surface suspicious patterns with more context.

Build notes

Started from the weakness of flat tabular fraud models and made relationships the main signal surface.

Designed the pipeline so the graph layer could support both modeling and downstream analyst workflows.

Focused on explainability because fraud systems fail when they score risk without usable context.

Tools + stack

Neo4j, Graph Data Science, Python, and analyst dashboard.

PythonPyTorchGraph MLFastAPIPostgreSQL

Core edge

Relational reasoning

Main surface

Fraud context

Analyst value

More explainable alerts

Architecture flow

Step 1

events

Step 2

graph build

Step 3

node features

Step 4

fraud model

Step 5

alerting

Why graph here

Fraud often hides in coordination. Graph features help reveal rings, mule links, device reuse, and indirect relationships that flat tabular models can miss.

System shape

The pipeline turns events into graph structure, extracts node and edge features, scores risk, and exposes suspicious activity through a service layer for downstream action.

Decision signals

Use relationships as signal, not decoration

Expose suspicious structure, not only scores

Design for offline modeling and alert operations

Outcomes

Used relationships as first-class signal instead of treating transactions as isolated rows.

Made fraud reasoning more explainable by showing the suspicious cluster structure.

Created a design that could support both offline modeling and analyst-facing alerting.