How Nubank Uses Transformers to Model Financial Habits at Scale

Nubank's AI Core team developed nuFormer, a Transformer-based foundation model that learns representations directly from raw financial transaction sequences using self-supervised learning. Rather than relying on hand-crafted features and traditional tree-based ML models, nuFormer treats financial transactions as tokenized sequences — analogous to natural language — to capture complex temporal behavioral patterns. The model supports multimodal inputs including structured data, behavioral signals, and in-app interactions, and can be fine-tuned for credit scoring, fraud detection, income prediction, and product recommendations. Deployed in production across 131 million customers, it has delivered gains in key metrics within months that surpassed years of incremental improvements from traditional approaches. Key engineering challenges include high-latency inference, distributed training infrastructure, continuous monitoring, and multi-team governance required for financial compliance.