Nubank: From LLMs to Financial Inclusion: Efficient LLM Training and Scaling AI Agents for 131 Million Lives

Nubank serves 131 million customers across Brazil, Mexico, and Colombia, many of whom accessed financial services for the first time through the platform. This scale demands AI that is both technically excellent and deeply human, capable of understanding individual behavior and delivering personalized, trustworthy experiences across languages and financial backgrounds. We present three interconnected lines of work across Nubank’s AI stack: 1. Efficient model training: research on more efficient optimization to enable training larger, more capable models under real-world resource constraints. 2. Deep learning from user behavior: nuFormer, a transformer-based self-supervised model that learns rich user representations directly from raw transaction sequences to improve large-scale recommendation systems. 3. Production AI agents for customer support: an evaluation-driven framework with a closed loop of context engineering, prompt optimization, LLM judges, and simulation-to-production validation to build high-quality AI agents that serve millions of customer interactions.