Finance ops · Production

Nubank deploys billion-parameter Foundation Models across predictive decision engines in first eight months after Hyperplane acquisition

The problem

Nubank historically relied on linear models, gradient-boosted trees, and aggregated tabular features for predictive AI decisions, which could not capture complex behavioral signals, limiting the bank's ability to advance to an AI-first model.

Workflow diagram · grounded in source
1
Hyperplane acquisition triggers project
trigger
“Nubank acquired Hyperplane in July 2024 to integrate our small team's Foundation Model technology into one of the world's largest digital banks quickly and effectively”
2
Sequence data ingestion and validation
integration
“We formed a dedicated team to ingest, validate, and enrich Nubank's enormous trove of rich transaction (and non-transaction) data sources”
3
Train billion-parameter Foundation Models
ai_action
“We leverage Ray to enable our small infra team to scale out heterogeneous clusters, allowing ML Engineers to train billion-parameter models on all 100+ million Nubank customers and their transaction histories. Many decisions at Nubank oc…”
4
Challenger vs baseline evaluation
validation
“we onboard product use cases and train Foundation Model-based challengers against Tabular ML baselines. Building a tight, reliable experiment iteration loop is paramount when introducing significant modeling changes to an existing produc…”
5
Deploy to production decision engines
output
“deployment of large-scale transformer-based sequence models in several key decision engines”
Reported outcome

Over the initial project period, Nubank achieved a +1.20% AUC lift across benchmark tasks—described as 2~3x the lift for a mature model's typical annual release—and deployed large-scale transformer-based sequence models to several key decision engines, all without adding any new data source.

Reported metrics
average AUC lift across benchmark tasks+1.20%
AUC lift relative to mature model annual release2~3x the lift for a mature model's typical annual release
customers covered by Foundation Models100+ million
Tokens processed during trainingO(trillions) tokens
Show all 5 reported metrics
average AUC lift across benchmark tasks+1.20%
AUC lift relative to mature model annual release2~3x the lift for a mature model's typical annual release
customers covered by Foundation Models100+ million
tokens processed during trainingO(trillions) tokens
lift achieved without new data sourceswithout adding any new data source
Reported stack
Ray
Source
https://building.nubank.com/foundation-models-ai-nubank-transformation/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Over the initial project period, Nubank achieved a +1.20% AUC lift across benchmark tasks—described as 2~3x the lift for a mature model's typical annual release—and deployed large-scale transformer-based sequence mode…

What tools did this team use?

Ray.

What results were reported?

average AUC lift across benchmark tasks: +1.20%; AUC lift relative to mature model annual release: 2~3x the lift for a mature model's typical annual release; customers covered by Foundation Models: 100+ million; Tokens processed during training: O(trillions) tokens (source-reported, not independently verified).

How is this finance ops AI workflow structured?

Hyperplane acquisition triggers project → Sequence data ingestion and validation → Train billion-parameter Foundation Models → Challenger vs baseline evaluation → Deploy to production decision engines.

Related finance ops cases