funding · November 6, 2026
By Joe Jurczyk, Editor
Guava raises $12.5M seed for voice agents in regulated industries
The Los Angeles developer platform, founded by alumni of Marqeta and voice lab Gridspace, raised a $12.5m seed led by Acrew Capital to run its own speech models in a single stack.
Guava, a Los Angeles-based developer platform for AI voice agents, has raised $12.5m in seed funding led by Acrew Capital, with Flourish Ventures, WillowTree Ventures and Clocktower Ventures participating alongside existing pre-seed investor Commerce Ventures. The company targets reliable automated conversations in regulated industries, including insurance, and builds and runs its own models in a single stack rather than stitching together third-party speech recognition, language and speech synthesis models.
Founders are Omri Dahan (CEO, formerly chief revenue officer at Marqeta), Anthony Scodary and Nico Benitez, both from the voice research lab Gridspace, and Mike Piechna. Gridspace's models were trained on more than 10 billion minutes of live conversations from regulated call centres. Guava's first public model, Daytona, launched in September; the company says more than a dozen enterprise customers process tens of thousands of calls a day on it, that Daytona has ranked in the top 10 on Coval's word error rate benchmark and often first or second, and that enterprise customers typically reach their first live regulated call within 14 days of starting implementation. The new money goes to models and product, an enterprise and BPO go-to-market team, and voice AI for edge devices, on-premises installations and full-duplex conversations.
Why it matters for digital humans: Any conversational avatar inherits the failure modes of the voice layer underneath it, and in regulated contexts those failures are compliance events, not bad demos. Guava's pitch — one vertically integrated stack instead of an ASR-LLM-TTS pipeline — is the same architectural argument now being made at the video layer by full-duplex avatar models, and for the same reason: cascades add latency and lose paralinguistic signal at every hop. The customer count, call volume and 14-day implementation figures are company-reported and unaudited; the Coval benchmark placement is at least externally measured.