Kumo, a leader in predictive AI, has launched KumoRFM-2, a groundbreaking foundation model designed for enterprise data. This new model is the first of its kind to outperform traditional supervised machine learning on complex relational databases. It promises to replace months of specialized data science work with a single, queryable model accessible to any team.
A New Paradigm for Enterprise Data
Traditional predictive models struggle with enterprise data, requiring that interconnected tables be flattened into a single file. This process destroys valuable relational signals, which are crucial for accurate predictions and demand extensive manual engineering. KumoRFM-2 addresses this by operating directly on the native graph structure of a database without flattening.
The model's unique approach preserves the intricate relationships between entities like customers, products, and transactions. This allows it to uncover deeper insights that are typically lost in conventional data preparation pipelines. As a result, businesses can generate predictions with zero feature engineering and no task-specific training, accelerating time-to-value.
Breakthrough Architecture and Performance
At its core, KumoRFM-2 is built on a novel Relational Graph Transformer architecture. This design enables the model to process data at an impressive 5 GB per second and scale to over 500 billion rows. The architecture allows the AI to attend to any data point, preserving the complete structure of relational data.
The model has demonstrated state-of-the-art results across 41 predictive tasks and four major benchmarks. On the SAP SALT enterprise benchmark, it surpasses tabular model ensembles like AutoGluon by a significant margin. It also outperforms the strongest supervised models on the Stanford RelBenchV1 benchmark, showcasing its superior accuracy.
Democratizing Predictive Analytics
KumoRFM-2 is designed for accessibility, featuring a natural-language interface that allows users to ask predictive questions in plain English. This capability empowers business users without data science expertise to generate sophisticated forecasts and insights instantly. The system translates these queries into Kumo's Predictive Query Language, making advanced AI accessible across an organization.
The real-world impact is already evident with clients like Databricks, which reported a significant improvement in lead conversion rates. The company saw conversions jump from 1.2x to 6x and doubled its volume of high-quality leads. This demonstrates the platform's ability to drive substantial marketing performance and tangible business outcomes.
The Vision of Industry Veterans
The development of KumoRFM-2 was led by a team of seasoned AI experts with backgrounds at Pinterest, Airbnb, and LinkedIn. Co-founders Dr. Vanja Josifovski, Dr. Jure Leskovec, and Dr. Hema Raghavan bring decades of experience in deploying AI at scale. The team is also credited with creating PyTorch Geometric, a widely used library for graph machine learning.
The company's innovative approach has attracted significant support from the technology community. Kumo is backed by venture capital firm Sequoia Capital and advised by industry leaders from Snowflake, Databricks, and Pinterest. This strong backing underscores the confidence in Kumo's mission to redefine predictive AI for the enterprise.
The launch of KumoRFM-2 marks a significant milestone in the evolution of enterprise AI. By eliminating the need for complex feature engineering and specialized training, Kumo is making predictive intelligence more efficient and accessible. This innovation positions the company to unlock immense untapped value from the structured data that powers global businesses.