Kumo's early team members and investors gather at a company dinner.
Modern businesses possess complicated networks of data, connecting information like customer behavior to marketing campaigns or fraud detection. But, to run useful AI predictions on the data often requires untangling the web of data connections. A new Stanford-bred startup says it has a solution using a new class of artificial intelligence to solve that problem.
Kumo announced itself to the world on Thursday with $18.5 million in Series A funding that it hopes will help it become the go-to software for AI prediction in the “modern data stack,” a set of cloud computing tools to store and harness large quantities of data. Sequoia Capital led the round at a valuation of $100 million; additional participation came from Ron Conway’s SV Angel and his son Ronny Conway’s A Capital.
The Mountain View, California-based startup was launched four months ago by founders Vanja Josifovski (formerly chief technology officer at Pinterest and Airbnb’s Homes business), Hema Raghavan (an ex-LinkedIn engineering director) and Stanford professor Jure Leskovec, who was also previously Pinterest’s chief scientist. The company comes as the culmination of five years of academic research conducted by a Stanford team featuring Leskovec, in conjunction with Germany’s Dortmund University. They focused on a budding form of AI, termed “graph neural networks,” which approaches machine learning by treating the data as if it were a complex graph network. Older forms of neural networks have become good at tasks with “structured data,” like image recognition or speech detection, but are hampered by data with unordered connections.
The research led to the development of PyG, an open source tool for graph neural network learning that was first launched five years ago. In the intervening time, Kumo’s founders implemented the technology at Pinterest and LinkedIn. “LinkedIn is like one big graph,” as Josifovski, the CEO, puts it, before contending that graph neural networks have “the potential to revolutionize machine learning in a similar way that deep learning revolutionized speech.”
But whereas large tech companies have the resources and manpower to develop these tools with in-house teams, most companies cannot do the same. That’s where Kumo comes in. The company’s software leverages the tech from PyG as the foundation for its software that helps customers to more easily craft complex predictive models from their business data. “Today, you can find out how many clients churned after 30 days,” Josifovski says. “Kumo is aiming to provide the same functionality for the future—the next 30 days.” Kumo’s product is designed primarily for data analysts and data scientists, and Josifovski says it should be usable even for employees without tech expertise. “Every company is having problems hiring data scientists,” he says. “If we’re able to package in a consumer-centric way, it will have a profound impact on the computing world.”
Kumo will use the capital it raised to scale up the product features and continue to focus on research and development. The startup currently employs more than 20 people, most of them engineers from the Stanford-Dortmund network with expertise in graph neural networks. But so far, the startup has not generated any meaningful revenue. The subscription-based product is in beta testing, being used by “select clients,” says Josifovski, though he will not share any names, nor does he have a time line for when the product will become commercially available. According to Konstantine Buhler, the Sequoia partner who led the financing, Kumo has been searching for customers among the public market’s largest enterprise companies. “There’s a sucking sound here,” he says. “The market wants this.”
Still, Kumo will have a tall task to bring graph neural networks into the mainstream. Companies valued in the billions of dollars, like Databricks, DataRobot and Dataiku, have already established lucrative businesses on different approaches to data science. Josifovski says Kumo is solving similar problems for some of those firms. “But, we intend to make machine learning an order of magnitude simpler,” he says. “We are fundamentally trying to leapfrog the current state of AI and render obsolete current methods.”
