Context
The problem
Fraud signals are scattered across payments, sessions and behaviour, and you only get a few milliseconds to decide.
Approach
How I built it
LSTM and autoencoder models score how unusual the behaviour looks across different kinds of signals.
The scoring is built to be fast, with a target of under 50ms per decision.
Monitoring was part of the plan from the start, for payment risk, bots and exam integrity.
Result
What came out of it
Decisions in under 50ms
One system for three kinds of fraud
Monitoring built in