.svg)
A top tier global payment processor needed to identify fraud in real-time across billions of transactions it manages each day. Effective fraud detection relies on accurate real-time processing of tens of thousands of transactions per second. As transaction volume and sophistication of fraud schemes grow over time, fraud prevention systems must become increasingly efficient to combat volume, while also increasing accuracy in fraud detection.
- High Operational Demands: Needed to process up to 65K transactions per second with sub-2ms latency per transaction while maintaining accuracy.
- Slower Model Iteration: Long training cycles limited how quickly new fraud detection strategies could be tested and deployed.
- Inefficient Precomputation: Prefill (CPU-side preprocessing) introduced latency in updating user states and rolling out new models, along with operational overhead and downtime.
- Massive State Storage Constraints: Each card’s transaction history produced a precomputed state that had to be saved for real-time decoding. With billions of active card sequences, even small increases in state size would multiply into petabytes of additional storage, making scalability and infrastructure cost a critical constraint.
We partnered closely with the client to develop a custom model powered by the Liquid Foundation Model (LFM) backbone, designed to deliver speed, scalability, and stronger detection performance.
- Efficient Training via Packing: Replaced GRU padding with numerically correct packing – eliminating wasted compute (up to 80%) and achieving faster training.
- Custom Embedding & Inference Stack: Optimized for CPU inference, maintaining below 1ms latency.
- Compact State Representation: Maintained per-user state size, ensuring no additional storage burden at billion-scale deployment.
- Privacy-Focused Deployment: Deployed fully on-premise / on-tenant infrastructure, aligning with the client’s security and compliance standards.
Higher High-Value Fraud Capture
- Detects a higher share of total fraud dollars, resulting in an estimated $230 million in additional fraud losses prevented annually
- Directly reduces financial losses, chargebacks, and customer disruption
- Outperformed legacy GRU model on custom dollar-weighted accuracy metric
Real-Time Fraud Detection at Scale
- Outperformed current real-time prediction latency by over 2x
- Seamless scalability across billions of card sequences
Faster Training & Development
- 1.8x faster training per token + 80% more tokens per batch via packing
- Enables rapid R&D iterations and faster deployment of new fraud strategies
Scalable & Efficient Infrastructure
- No increase in precomputed user state size or storage footprint
- CPU-optimized inference stack with on-premise deployment for privacy compliance
Enhanced Customer Trust & Protection
- Real-time fraud detection provides immediate financial protection and prevents costly, time-consuming recovery efforts for users.
- Customers experience greater confidence and trust in the payment processor (the client), with reduced fraud-related stress and protection of their credit health.