There’s a chance that your next phone call or video conference is answered by an AI model. Voice and video models have quickly grown capable of powering real time, interactive experiences that feel lifelike. Behind the scenes of these experiences, however, engineers have the daunting task of keeping the lights on by managing and scaling the complex set of computational resources these models use.

We were thrilled to meet Michael Louis and Jono Irwin who had both lived through these challenges as they scaled AI products at their prior startup OneCart which was acquired by Walmart-owned Massmart. At OneCart the team’s workflow mirrored that of many developers – it involved stitching together a wide range of tools for experimentation, deployment, and debugging AI models. Even after successfully implementing their solution, costs were high and maintaining the sprawling array of infrastructure products was problematic. Michael and Jono founded Cerebrium to set out to build the infrastructure they wished they had.

When we tried out Cerebrium’s platform which was built by a surprisingly small, focused team of engineers, a few things stood out to us. First, the platform is remarkably fast with cold start times of 2-4 seconds and network latency of less than 50ms. What enables this speed is a custom container runtime, caching system, and scheduler which dispatches inference close to client machines. Every millisecond of latency shaved off results in less compute billed to developers and Cerebrium’s customers spend 40% less on compute than they did previously. Secondly, the developer experience is seamless and allows users to easily experiment with models before deploying them. Developers can build containers in 10s of seconds and easily dictate resource types or allocations within a config file. Lastly, Cerebrium is obsessed with reliability and works hard to maintain five nines of uptime with a team across all timezones to support developers.

Cerebrium powers some of the most innovative voice and video AI companies including Tavus, Deepgram, and Vapi. Hundreds of millions of seconds of computation are run on servers Cerebrium orchestrates every month. Their platform enables small teams to build captivating experiences and we encourage you to try out their technology here.

We are pleased to share that we’ve led Cerebrium’s $8.5M seed round with participation from Y Combinator, Authentic Ventures, and several strategic angels and operators.