Source: Forbes — 2026-07-08
Summary
Revolut, working with Nvidia, built PRAGMA, a family of transformer-based foundation models trained on roughly 40 billion banking events from about 25 million users across 111 countries (207 billion tokens), to handle fraud detection, credit scoring, and product recommendations through one shared backbone instead of separate task-specific models. On Revolut's internal benchmarks, PRAGMA delivers a 64.7% improvement in fraud recall, 2.3x better accuracy on credit-default risk, and 41% more relevant product recommendations versus the prior production models it replaced.
Key Takeaways
- The architectural bet: instead of maintaining separate specialized models for fraud detection, credit scoring, and recommendations, PRAGMA is one transformer backbone fine-tuned across all of them — so an improvement to the shared backbone flows through to every downstream task at once, rather than requiring each team to retrain its own model.
- Training scale: ~40 billion banking events from ~25 million users in 111 countries, totaling 207 billion tokens — the largest variant, PRAGMA-L (1B parameters), was trained using up to 64 Nvidia H100 GPUs on Nebius AI Cloud infrastructure.
- Verified performance deltas against the prior production models: 64.7% improvement in fraud recall (with a reported 17% precision gain alongside it), 2.3x better accuracy at flagging credit-default risk, and 41% more relevant product recommendations.
- This is a case study in domain-specific foundation models beating a "call a frontier general-purpose LLM API for everything" approach for a company with enough proprietary transaction data to train its own backbone — a path only available to institutions with Revolut's data scale.
Reel Script
Hook (16s / 36 words) Revolut didn't bolt a chatbot onto its banking app — it trained its own foundation model on 40 billion of its own transactions, and it's now catching 65% more fraud than the system it replaced.
Core Concept (70s / 155 words) Most banks run fraud detection, credit scoring, and product recommendations as three completely separate systems — different models, different teams, different training pipelines, none of them learning from each other. Revolut's approach with PRAGMA is different: it's one transformer-based foundation model — the same underlying architecture behind large language models, but trained on banking events instead of text — that gets fine-tuned to handle all three tasks. Think of it like this: instead of three separate specialists who each learn banking from scratch, you train one generalist that deeply understands the raw shape of a customer's financial behavior, then teach it three different specializations on top of that shared understanding. The advantage is that when Revolut improves the shared backbone — say, giving it a better sense of what "normal" spending looks like — every one of those three tasks gets better simultaneously, instead of three separate teams each having to independently retrain their own model.
Hands-On (85s / 190 words) The training data itself is the artifact worth dwelling on: roughly 40 billion individual banking events, pulled from about 25 million Revolut users across 111 countries, totaling 207 billion tokens — for scale, that's a dataset built entirely from real transaction-level behavior, not scraped text. The largest model in the family, PRAGMA-L, sits at 1 billion parameters and was trained using up to 64 Nvidia H100 GPUs on Nebius's AI cloud. On the results side, Revolut reports these deltas against its previous production models, not against some abstract baseline: a 64.7% improvement in fraud recall — meaning it catches significantly more actual fraud cases than before — while also improving precision by 17%, so it's not just casting a wider net and creating more false alarms. Separately, credit-default risk prediction improved 2.3x in accuracy, and product recommendations came back 41% more relevant. Three different business functions, one shared model, three measured improvements.
Takeaway (24s / 53 words) This only works if you have Revolut-scale proprietary transaction data to train on — but it's a real signal that domain-specific foundation models, not general-purpose LLM APIs, are where serious fintech infrastructure is heading next. Worth tracking as a template, even if you can't replicate the data scale yet.