Financial Services
NeoSpace Develops Large Tabular Foundation Models for Financial Institutions
Objective
NeoSpace, an NVIDIA Inception member, helps enterprises turn trillions of tabular, event, and unstructured records into real-time predictions and decisions using enterprise large tabular foundation models. Its NeoData platform, powered by NVIDIA accelerated computing, delivers the performance and operational scalability required for complex enterprise data processing. Banks, telecom operators, insurers, and companies across multiple industries are already embracing this new technology to unlock greater intelligence and value from their data. One of Latin America’s largest private banks used NeoSpace’s foundation models to significantly improve credit decisioning and unlock installment credit offers.
Customer
NeoSpace
Partner
Oracle Cloud Infrastructure (OCI)
Topic
Generative AI / LLMs
Products
Key Takeaways
Higher Foundation Model Accuracy
- 30%–50% higher model accuracy compared with traditional ML approaches, delivering more consistent predictions across populations, data modalities, and time.
Fine-Tuned Model Performance
- 19% lift over NeoSpace’s baseline model, due to more training iterations enabled by NVIDIA AI infrastructure.
Improved Credit Accessibility
- 10% increase in installment credit offers, driven by higher hiring and conversion rates in the top decile of NeoSpace’s customers.
Scaling Transformer Models for Real‑Time Decisions
Before utilizing NVIDIA’s accelerated computing platform, NeoSpace saw enterprise workloads built to run on traditional ML models struggled to keep pace with data volume and complexity.
As its customers moved to transformer-based foundation models trained on enterprise datasets, bottlenecks appeared in time-to-train as well as inference throughput, latency, and costs. For large financial institutions these constraints translated into slower experimentation, limited use cases, and difficulty delivering real-time, personalized decisions across tens of millions of end customers.
NeoSpace needed a way to make foundation models in financial services production-grade and economically viable, so banks could scale AI applications on top of trillions of enterprise records.
NeoData training runs a dashboard for domain-specific models, including investment strategies.
NeoData on NVIDIA Blackwell With CUDA
NeoSpace architected its NeoData platform to run on NVIDIA accelerated computing for both training and inference—transforming its customers’ structured and unstructured data into real-time, actionable intelligence.
NeoData now operates at scale across a multi-cloud environment for one of Latin America’s largest private banks with more than 60 million customers.
NeoSpace is among the first companies in Latin America to deploy NVIDIA GB200 NVL72 systems in production, through Oracle Cloud Infrastructure (OCI).
NeoSpace’s stack is also built on NVIDIA® CUDA® libraries to accelerate training and serving of its foundation models, including:
- Core GPU Programming: CUDA Toolkit, NCCL, NVSHMEM, GPUDirect®
- Numerical Libraries and Primitives: cuBLAS, CUTLASS, CuTe
- AI Frameworks and Data Science: RAPIDS™, NVIDIA TensorRT™ LLM, Dask
NeoSpace orchestrates its NVIDIA GPUs using NeoCore, an internal tool that runs directly on bare-metal hosts in its multi-cloud environments to manage workload distribution. This architecture allows NeoSpace to run foundation models as a predictive layer across use cases—replacing fragmented ML pipelines with a unified, production-grade approach.
NeoData inference server dashboard monitors system health, performance, and resource usage for specific model checkpoints across investments, risk assessment, volatility predictions, and more.
Higher Accuracy and a Lift in Installment Credit Offers
Compared with traditional ML models, NeoSpace’s foundation models powered by NVIDIA delivered improved results across customer populations, data modalities, and time. NeoSpace achieved at least 30%–50% higher accuracy, enabling its financial services customers to expand AI use cases without re-architecting its infrastructure.
In a flagship deployment, NeoSpace worked with another major Latin American financial institution to build a propensity modeling solution on large-scale tabular data to identify customers most likely to contract installment credit.
The model was evaluated using Kolmogorov–Smirnov (KS), hiring rate, and conversion rate across deciles. This deployment impacted 6 million customers in production and leveraged more than 10 TB of data covering over 50 million customers including CRM feature stores, eligibility variables, historical contracting behavior, transactional data, and more.
Against the bank’s existing baseline, NeoSpace delivered a 10 to 20 percentage point improvement in KS. The hiring rate in the top decile increased by 20 to 30 percentage points, showing that a larger share of customers who contracted the bank’s installment credit offer were concentrated in the highest-propensity segment. These performance gains enabled the bank to contact fewer customers while achieving higher overall results, increasing efficiency per interaction and enabling more high-impact campaigns.
Overall, the improved propensity strategy translated into a 10% uplift in installment credit offers, driven by better ranking of high-propensity customers and precise engagement. These foundation models are the institution’s new core predictive layer that delivers more stable predictions, faster time-to-value, and improvement in business efficiency.
The Next Phase With NeoSpace
NeoSpace is testing a fused model that combines the current tabular model with event-based transactional data, such as credit card transactions, checking account activity, and bill payments, and is expected to add roughly 5 percentage points of additional installment credit offer lift.
The bank is already shadowing this propensity modeling framework from a single product to nearly 50 financial products, creating a platform that understands each customer, anticipates needs, and scales personalized engagement.
By building NeoData on the NVIDIA Blackwell platform and CUDA software stack, NeoSpace deployed production-grade financial services foundation models that link data to decisions for some of Latin America’s largest financial institutions.
“Enterprise foundation models are the missing layer between data and decisions. With NVIDIA, we’re making that layer production-grade, training and serving at scale, so customers can unlock predictive value from trillions of enterprise records.”
Bruno Pierobon
CEO, NeoSpace
Explore leading financial services companies using AI to optimize trading, detect transaction fraud, and enhance customer experiences.