Ledgerly
A real-time machine-learning pipeline that scores every transaction across three payment gateways in under 80 milliseconds, without slowing checkout.
Numbers that moved
The challenge
Ledgerly's existing rules engine caught obvious fraud but missed sophisticated patterns, while flagging enough good customers to hurt conversion.
They needed a model that improved precision without adding latency across three separate payment gateways with different data shapes.
Our approach
We built an evaluation set from a year of labeled transactions before touching production, so every model change was measured against a fixed bar.
A feature pipeline unified signals across all three gateways into one real-time scoring path, designed to fail open rather than block a good transaction.
What we built
A streaming feature store and gradient-boosted model serve scores in under 80ms, with a shadow-mode rollout that proved the model before it made a single live decision.
A reviewer console gives the fraud team full explainability on every flagged transaction, with a feedback loop that retrains the model monthly.
We went from arguing about false positives every week to trusting the system. Chargebacks dropped and our support team stopped hearing complaints about wrongly blocked cards.
What we used
Related work
Need fraud detection that doesn't cost you conversions?
We build real-time scoring pipelines tuned for precision, latency, and explainability together.