Graph-Based Fraud Detection — Loan Origination

Situation & Challenge

A microfinance institution was experiencing material portfolio losses from ghost borrower fraud and loan stacking. The fraud was invisible to individual application review because patterns only emerged at network level — across borrowers, addresses, phone clusters, and identity documents shared across multiple applicants.

What Coadjutant Built

A graph neural network model resolving hidden entity relationships across seven data sources — national identity, phone networks, address clusters, bureau records, and repayment history. Fraud signals scored at origination in under 400ms. Designed with a low false-positive rate to minimise friction for legitimate borrowers while blocking fraud rings at first application attempt.

Technology & Protocols