Sherpa, a startup from Bilbao, Spain that was an early mover in building a voice-based digital assistant and predictive search for Spanish-speaking audiences, has raised some more funding to double down on a newer focus for the startup: building out privacy-first AI services for enterprise customers.
The company has closed $8.5 million, funding that Xabi Uribe-Etxebarria, Sherpa’s founder and CEO, said it will be using to continue building out a privacy-focused machine learning platform based on a federated learning model alongside its existing conversational AI and search services. Early users of the service have included the Spanish public health services, which were using the platform to analyse information about COVID-19 cases to predict demand and capacity in emergency rooms around the country.
The funding is coming from Marcelo Gigliani, a managing partner at Apax Digital; Alex Cruz, the chairman of British Airways; and Spanish investment firms Mundi Ventures and Ekarpen. The funding is an extension to the $15 million Sherpa has already raised in a Series A. From what I understand, Sherpa is currently also raising a larger Series B.
The turn to building and commercializing federated learning services comes at a time when the conversational AI business found itself stalling.
Sherpa saw some early traction for its Spanish voice assistant, which first emerged at a time when efforts from Apple in the form of Siri, Amazon in the form of Alexa, and others hadn’t really made strong advances to address markets outside of those where English is spoken.
The service passed 5 million users as of 2019 — customers using its conversational AI and predictive search services include the Spanish media company Prisa, Volkswagen, Porsche and Samsung.
But as Uribe-Etxebarria describes it, while that assistant business is still chugging along, he came up against a difficult truth: the biggest players in English voice assistants eventually did add Spanish, and the conversational AI investments they would make over time would make it impossible for Sherpa to keep up in that market longer-term on its own.
“Unless we did a big deal with a company, we wouldn’t be able to compete against Amazon, Apple and others,” he said.
That led the company to start exploring other ways of applying its AI engine.
It came on to federated privacy, Uribe-Etxebarria said, when it started to look at how it might expand its predictive search services into productivity applications.
“A perfect assistant would be able to read emails and know which actions to take, but there are privacy issues around how to make that work,” Uribe-Etxebarria said. Someone suggested to him to look at federated learning as one way to “teach” its assistant to work with email. “We thought, if we put 20 people to work, we could build something to read and respond to emails.”
The platform that Sherpa built, Uribe-Etxebarria said, worked better than they had anticipated, and so a year later, the team decided that it could use it for more than just triaging email: it could be productized and sold to others as an engine for training machine learning models with more sensitive data in a more privacy-compliant way.
It’s not the only company pursuing this approach: TensorFlow from Google also uses federated learning, as does Fate (which includes cloud computing security experts from Tencent contributing to it), and PySyft, a federated learning open-source library.
Sherpa is working with several companies under NDAs in areas like healthcare, and Uribe-Etxebarria said it plans to announce customers in other areas like telecoms, retail and insurance in the near future.