How Nubank Uses AI Models to Analyze Transaction Data for 100M Users

Nubank processes transaction data from 100 million users using transformer-based foundation models instead of traditional manual feature engineering. Their system converts raw transactions into tokenized sequences, trains models using self-supervised learning on trillions of transactions, and combines sequential embeddings with tabular data through joint fusion architecture. The centralized AI platform allows teams across the company to access pretrained models for various financial tasks like credit scoring, fraud detection, and personalization.