Revolut built PRAGMA, a family of encoder-style transformer foundation models on NVIDIA H100 GPUs (Nebius cloud). Pre-trained on ~26M user records, 24–40B events, 207B tokens across 111 countries over ~28 months.
PRAGMA-S (10M params) converges in ~2 days on 16 GPUs; larger variants ~2 weeks on 16–32 GPUs. Sub-second latency for real-time fraud screening.
Internal benchmarks vs prior production models: 64.7% fraud recall improvement, 16.7% fraud precision lift, 2.3× credit default risk accuracy, 41% more relevant product recommendations.
Replaces siloed task-specific ML across fraud, credit, engagement, and recommendations with one shared behavioral intelligence layer.