Fraud Detection · Marketing Personalization · Risk Management
Revolut needed to replace siloed fraud, credit, marketing, and recommendation models that depended on months of feature engineering with a single behavioral foundation model that could learn from raw financial event streams and scale efficiently across fast-growing markets.
Reported impact
Revolut reports 3–5x faster model-development cycles, 2–5x higher training throughput, 2.3x better credit default risk accuracy, 65 percent more fraud cases caught with 17 percent higher fraud precision, and 41 percent more relevant product recommendations by consolidating on PRAGMA.
Revolut rebuilt its AI stack around PRAGMA, a transformer-based family of behavioral foundation models trained on NVIDIA accelerated computing. The architecture uses three encoders for user attributes, individual events, and long-term history, then applies structured tokenization to convert profile fields, transaction streams, amounts, and timestamps into a shared temporal representation. NVIDIA Llama-Nemotron-Embed-1B-v2 enriches the backbone with embeddings from unstructured merchant descriptions. Revolut pre-trained PRAGMA on NVIDIA H100 GPU clusters in Nebius AI Cloud, using shard-based dynamic batching, sequence packing, and variable-length attention to raise throughput and fit training into fixed windows. Downstream teams either freeze the backbone and add lightweight heads or apply LoRA fine-tuning, so fraud, credit, recommendations, and growth models all reuse the same behavioral intelligence layer.
Raw customer attributes, transactions, timestamps, and merchant text are tokenized into PRAGMA, H100 clusters pre-train shared embeddings on Nebius AI Cloud, and downstream fraud, credit, recommendation, and growth tasks consume the shared backbone through linear heads or LoRA adapters.
Major gains over production baselines, including 2.3× better accuracy at identifying credit default risk, 65% more fraud cases caught while also being 17% more precise in fraud alerts, and 41% more relevant product recommendations.
NVIDIA Llama‑Nemotron‑Embed‑1B‑v2 is used to embed unstructured text fields such as merchant descriptions, enriching the behavioral signal and delivering a measured 16.1% improvement in credit risk prediction performance.
Revolut is replacing multiple siloed systems across fraud, credit, engagement, and recommendations with one shared behavioral intelligence layer.
Teams can either freeze the model and train lightweight linear heads on top of embeddings for rapid experiments or apply LoRA fine-tuning, updating only 2%–4% of parameters.
With NVIDIA H100 GPUs on Nebius AI Cloud via sequence packing and dynamic batching, making large-scale pre-training feasible within fixed windows.
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
Databricks Sample May 15, 2025 Maya Maya needed a unified real-time lakehouse because legacy architecture limited the scale of underwriting, fraud analytics, and personalized customer experiences across its financial ecosystem. $2.1 billion Databricks Assistant · Databricks Data + AI Platform · Delta Lake Databricks Sample Feb 5, 2025 Techcombank Techcombank needed to break down data silos and strengthen governance because fragmented banking data was slowing innovation in customer interaction, fraud prevention, and risk management. Ushering personalized banking for millions of customers 1,000+ Databricks Data + AI Platform · Databricks SQL · Delta Lake 
NVIDIA Sample Mar 17, 2026 Mastercard Mastercard wants to turn its transaction data into a reusable AI insights engine for payments and commerce, especially to strengthen cybersecurity and fraud detection, improve personalization and loyalty use cases, and… billions of transactions Databricks · Nvidia's advanced accelerated computing platform · large tabular model (LTM) 