Real-time fraud detection for a challenger payments rail

The client is a challenger payments rail processing high volumes of card-not-present transactions, where fraud losses and false declines both hit the business directly. Block too little and losses mount; block too much and good customers are turned away and churn.

As a regulated payments business, every automated decision has to be explainable for disputes and audit. A model that was accurate but opaque would have been unusable however good its raw numbers looked.

Card-not-present fraud had to be caught inline, in the few milliseconds before a payment is authorised, without adding latency that would slow or block good transactions. Speed and accuracy were in direct tension.

Every decision, an approval or a block, had to be explainable after the fact, so the fraud and dispute teams could stand behind it to a customer or a regulator. Blocking a genuine customer with no defensible reason was as costly as missing a fraud.

Off-the-shelf rules could not keep pace with evolving fraud patterns without also generating false declines that eroded good revenue. The rail needed detection that was both sharp and accountable.

We built a streaming engine that scores every transaction inline in around 38 milliseconds, fast enough to sit in the authorisation path without customers ever feeling it. Detection happens before money moves, not in a nightly batch after the loss.

Every model decision is logged with the factors behind it, so the fraud and dispute teams can explain any approval or block on demand. Explainability was treated as a first-class requirement, not a reporting afterthought.

We validated the model against historical fraud outcomes and ran it alongside the incumbent controls before it took over, so its accuracy and its false-decline rate were both proven on real traffic before it owned the decision.