Revolut has developed a transaction foundation model called PRAGMA, utilizing NVIDIA’s accelerated computing platform, to enhance fraud detection, product recommendations, and other financial service applications, achieving 2.3 times better credit risk accuracy and up to five times higher training throughput. Founded in 2015, Revolut serves over 70 million customers across 40 markets and aims to streamline its operations by learning rich behavioral insights from raw event data. The implementation of PRAGMA allows for faster model development cycles—reportedly 3 to 5 times quicker—thanks to shared embeddings that eliminate the need for repetitive feature engineering, aligning with a broader industry trend towards adopting foundation models for improved efficiency in financial services.

NVIDIA: NVIDIA develops graphics processing units and full-stack AI platforms that power accelerated computing workloads. Its H100 GPUs and software stack enabled Revolut to pre-train and optimize the PRAGMA model on Nebius AI Cloud with techniques such as dynamic batching and variable-length attention for efficient handling of large-scale financial event sequences.
PRAGMA: PRAGMA is a family of transformer-based behavioral models created by Revolut that processes customer transaction streams, attributes, and history into unified embeddings for financial services. It functions as a reusable backbone for applications including fraud detection, credit risk assessment, marketing, and product recommendations, with configurations ranging from smaller efficient variants to larger high-accuracy models.
Revolut: Revolut is a London-based financial technology company offering banking, payments, foreign exchange, credit, and wealth management services. It developed PRAGMA, a family of transformer-based behavioral models, using NVIDIA’s accelerated computing platform to create unified representations from raw transaction data for multiple downstream financial tasks.

AI in Finance: Financial services firms are increasingly adopting foundation models trained on transaction data to replace fragmented, task-specific machine learning systems with shared behavioral representations.
Infrastructure Trends: Accelerated computing platforms from specialized hardware providers are being integrated into banking environments to support scalable training and low-latency inference of large transformer models on sensitive financial datasets.