## Overview Nubank, one of the world's largest digital banks, embarked on an ambitious initiative to integrate foundation models into their production AI platform to advance their "AI-First" vision. This case study represents a significant departure from traditional financial industry machine learning approaches, which typically rely on linear models, gradient-boosted trees, and aggregated tabular features. The initiative was accelerated through Nubank's acquisition of Hyperplane in July 2024, bringing together foundation model expertise with Nubank's decade-long experience in productionizing AI innovations in highly regulated financial environments. The core challenge addressed was enhancing predictive modeling capabilities across critical banking decisions by leveraging large-scale transformer architectures to process sequential customer data. Rather than completely overhauling their existing systems, Nubank adopted a pragmatic approach that preserved their proven data infrastructure, model governance, and deployment frameworks while introducing new capabilities for handling foundation models. ## Technical Architecture and Implementation Nubank's foundation model implementation centers on three key components that were developed from scratch while maintaining interfaces with existing systems. The sequence data preprocessing component transforms raw sequential data into formats suitable for transformer models, moving beyond the traditional ETL pipelines designed for tabular feature extraction. This represents a fundamental shift in how financial institutions handle customer behavioral data, processing entire transaction sequences rather than aggregated statistical summaries. The GPU and heterogeneous cluster infrastructure component addresses the computational requirements of billion-parameter models. Using Ray as their primary scaling framework, they built capabilities to train models on data from all 100+ million Nubank customers and their complete transaction histories. The scale is remarkable - their models process trillions of tokens during training, handling billions of labels with thousands of transactions per label for monthly decision cycles. The transformers and deep neural networks component implements the core modeling architectures. These billion-parameter models are designed to capture behavioral signals that traditional aggregated feature approaches cannot detect. The models are built to understand customer behavior patterns across diverse financial products and geographic markets, reflecting Nubank's multi-country digital banking presence. ## Production Integration Strategy A critical aspect of Nubank's approach is their careful integration strategy that minimizes disruption to existing production systems. They established clear interfaces between new foundation model components and existing AI platform layers, ensuring that proven model governance, data validation, and deployment processes remain intact. This design philosophy recognizes that financial institutions cannot afford to compromise on reliability and regulatory compliance when introducing new modeling approaches. The foundation model projects follow a structured challenger approach, where new models are trained and evaluated against existing tabular ML baselines using identical data sources and evaluation metrics. This methodology ensures fair comparison and helps quantify the incremental value of foundation models. Initially, the foundation models consume the same raw sequence data sources that feed into baseline models, but process them as sequences rather than aggregated features. ## Data Infrastructure and Monitoring The sequence data processing infrastructure represents a significant engineering undertaking. Beyond ingesting and transforming data, the team developed specialized tooling for validating and monitoring sequence data quality in production environments. This includes additional safeguards to prevent data leakage and ensure that sequential patterns don't introduce temporal biases that could compromise model performance. The data team tracks progress across three key stages: ingested data sources ready for model consumption, modeled data sources that have passed quality checks and demonstrated impact, and productionized data sources with full monitoring infrastructure. While initial implementations focus on transaction data, the architecture is designed to incorporate app events and product usage signals as additional sequence sources. ## Model Development and Evaluation Tools Nubank developed comprehensive tooling to support large-scale foundation model development across distributed teams. The Model Catalog provides centralized tracking of all trained models with comparison capabilities and filtering based on specific criteria. Each model entry includes complete artifacts, input data specifications, inference outputs, and training parameters, enabling systematic model management across multiple use cases. The Reporting Tool offers drag-and-drop analytics capabilities for rapid model evaluation and comparison. This tooling addresses the complexity of coordinating development across dozens of researchers and engineers working on horizontally deployable foundation models, ensuring unified progress toward common objectives. ## Performance Results and Business Impact The quantitative results demonstrate substantial improvements over traditional approaches. Across four benchmark tasks, foundation models achieved an average 1.20% AUC lift, which represents 2-3 times the typical annual improvement for mature models in financial services. Importantly, these improvements were achieved without incorporating additional data sources beyond those already used in baseline models, indicating that the foundation models are extracting previously inaccessible behavioral signals from existing data. The adoption metrics reveal systematic progress across critical use cases. The team tracks four deployment stages: problems onboarded with aligned evaluations, baselines replicated to ensure fair comparison, challengers built with sufficient metric improvements, and models deployed to production. This structured approach ensures that foundation models meet both performance and operational requirements before serving customers. ## Operational Considerations and Challenges While the blog post presents positive results, several operational challenges emerge from the implementation. The computational requirements for billion-parameter models processing trillions of tokens represent significant infrastructure costs that must be justified by business value. The complexity of coordinating development across multiple teams and ensuring model governance in regulated environments adds operational overhead compared to traditional ML approaches. The reliance on Ray for scaling heterogeneous clusters introduces technology dependencies that may limit flexibility or require specialized expertise for maintenance. Additionally, the focus on replicating existing baseline performance before deploying challengers, while methodologically sound, may slow innovation cycles compared to more aggressive deployment strategies. ## Strategic Implications for Financial AI Nubank's foundation model initiative represents a significant strategic bet on the transformative potential of large-scale AI in financial services. The emphasis on behavioral understanding through sequence modeling aligns with broader trends in AI research while addressing specific challenges in financial customer modeling. The acquisition of Hyperplane demonstrates commitment to building in-house AI capabilities rather than relying solely on external providers. The approach of preserving existing governance and deployment infrastructure while introducing new modeling capabilities offers a template for other financial institutions considering similar transformations. However, the substantial investment in specialized infrastructure, tooling, and expertise may limit applicability to smaller institutions without comparable resources. ## Future Directions and Scalability The roadmap focuses on three key areas: ingesting more unstructured data sources, training more advanced foundation models, and deploying to more critical use cases. The planned incorporation of app events and product usage signals could further enhance behavioral understanding, while continued architectural improvements may unlock additional performance gains. The success of this initiative positions Nubank to influence broader adoption of foundation models in financial services, potentially establishing new industry standards for AI-powered banking decisions. However, long-term success will depend on demonstrating sustained business value that justifies the significant operational complexity and computational costs associated with foundation model deployment at scale.