Revolut has launched Revolut Research, an AI division building its PRAGMA foundation model in collaboration with NVIDIA.
The new unit sits within the company's broader AI Department and is intended to provide a shared framework for the company's machine learning initiatives, rather than relying on separate, task-specific models.
PRAGMA and the NVIDIA collaboration
Revolut Research operates as the unit behind PRAGMA, a proprietary foundation model developed in collaboration with NVIDIA. According to the company, PRAGMA is designed to interpret financial behaviour patterns and to support real-time risk assessment, platform operations, and product recommendations across Revolut's services. The stated approach consolidates functions that might otherwise be handled by multiple separate models into a single underlying system.
Revolut said early testing of PRAGMA against historical data showed gains over the models it currently uses in production. The company reported 2.3 times higher accuracy in identifying credit default risk, a 65% increase in fraud cases detected alongside a 17% improvement in alert precision, and 41% more relevant product recommendations across retail and business accounts. These figures relate to internal comparisons against Revolut's existing systems rather than to external benchmarks or third-party verification.
Data scale and existing tools
Revolut's stated rationale for the initiative is the scale of data generated across its customer base, which the company said processes billions of cross-border transactions. The firm noted that its existing security models already review close to one billion transactions each month to identify fraudulent activity. In addition, Revolut Research's output is also intended to support front-end tools such as AIR, the company's in-app assistant designed to complete multi-step financial tasks, which is currently available only in the UK.
Pavel Nesterov, Head of AI at Revolut, said the unit was established to develop financial intelligence internally rather than through third-party or bolt-on software layers, framing it as consistent with an existing internal engineering approach. Revolut Research has already presented findings at academic and industry conferences this year, including ICML in Seoul. Furthermore, the company said the unit plans to publish research findings and open-source technical frameworks going forward, and will participate in NVIDIA's GTC Berlin event in October 2026 and ICAIF in November 2026, alongside quarterly meet-ups for the scientific community at Revolut's offices.
The launch reflects a broader trend among large fintech and banking players to develop proprietary AI infrastructure rather than integrate third-party tools, as firms compete on the ability to process and act on large volumes of customer transaction data in real time.
