from the editor

Tavus builds the infrastructure for real-time AI humans: video agents that see, hear, and hold conversations with sub-500 millisecond latency. The San Francisco company serves over 100,000 developers and enterprises including Deloitte and Amazon, deploying AI across healthcare, recruiting, education, and customer service.

The company builds three proprietary model families in-house: Phoenix for facial rendering, Raven for multimodal perception, and Sparrow for conversational timing and emotional understanding. That architecture sets Tavus apart from competitors like Synthesia, which license or aggregate third-party models rather than develop the underlying research.

The company's argument is that face-to-face video is the missing interface layer in AI, and that owning the full model stack is the only defensible way to hold that position. Whether that holds depends on whether proprietary rendering and perception can outpace commoditization as the field matures.