The new DNA of Fintech: how we’re redefining banking with Generative Intelligence
By: @AntonRepushko
The dividing line in banking won’t be between firms that use AI and firms that don’t. It’ll be between firms that build and firms that bolt on.
Bolting on means renting someone else's tools. Building means going deeper. That’s why we’re developing PRAGMA — our own foundation model, trained on billions of financial events rather than generic text.
One backbone that understands behaviour, so every team builds on the same intelligence. In our initial benchmarks we found:
• 65% more fraud caught
• 2.3x better credit risk prediction
• Models built in days instead of months
That’s the difference. Bolt on, and each problem needs its own tool. Build, and everything improves at once.
The Challenge: Siloed Models & Slow Experiments
Before PRAGMA, we followed the traditional pattern: dedicated ML pipelines for every single task, from fraud detection to marketing.
Each model depended on its own hand-crafted features and ETLs. Launching a new use case or market meant months of repetitive feature engineering.
This caused slow experimentation, fragmented decisions, and scaling limits. Fraud and credit systems worked from different customer representations, making holistic optimization incredibly difficult.
The Solution: One Unified Behavioral Backbone
Our goal with PRAGMA was to learn rich behavioral representations directly from raw event streams, then reuse those embeddings across risk, growth, and product workflows.
To achieve this at foundation-model scale, we’re working with NVIDIA and a mature deep learning software platform.
PRAGMA interprets each customer’s financial journey as a rich temporal signal, similar to how advanced LLMs interpret sequences of text.
Available in configurations from tens of millions to 1B parameters, it enables everything from ultra-efficient, low-latency inference to large-scale maximum predictive accuracy.
How We Trained the Model
We pre-trained PRAGMA on a massive dataset:
• ~26 million user records across 111 countries
24 billion events
• 207 billion tokens over ~28 months of history
Once pre-trained, PRAGMA serves as a shared backbone. Teams can freeze the model and train lightweight linear heads for rapid experiments, or apply LoRA fine-tuning (updating just 2%–4% of parameters).
Fraud, credit, marketing, and product teams can now spin up new models in days instead of months, without building new feature pipelines.
The Results
By consolidating on a PRAGMA backbone, we’ve achieved massive performance and efficiency lifts:
• 3–5× faster model development cycles
• 64.7% improvement in fraud recall over our prior production model
• 16.7% lift in fraud precision over our prior production model
Every update to the backbone flows through to multiple business units, compounding AI quality across the company. Our teams benefit from less duplicated code, simpler governance, and unified monitoring, allowing us to enter new markets without rebuilding our ML stack.
What's Next?
We plan to expand PRAGMA with:
• More multimodal inputs
• Continual pre-training to learn from new events continuously
• Broader downstream use cases (lifetime value prediction, churn forecasting, anomaly detection)
• A model-driven app interface that adapts in real time to user behavior
• Evolving into an AutoDL platform with embedding versioning and a full inference pipeline
Source: PRAGMA: Revolut Foundation Model, arXiv:2604.08649, CC BY 4.0.