← Back to BlogJune 2026Part of 10 Minutes or Less
Tavus CEO Hassaan Raza on building AI humans, not avatars (10 Minutes or Less)
Hassaan Raza, co-founder and CEO of Tavus, joined me on this week's podcast. Most podcasts run an hour or more. This one is 10 Minutes or Less.
Tavus is a research lab trying to make talking to a machine feel like talking to a real person. Founded in 2020, the company most recently raised a $40M Series B led by CRV, with customers including Salesforce, Amazon, and Alibaba.
What we covered in the rapid fire:
- "We're the human computing company." Hassaan rejects being lumped in with Synthesia, HeyGen, Hedra, or Google's Gemini. The difference: generating videos versus building a real-time AI human that sees you, hears you, reads your expressions, and reacts with EQ. Rendering, he says, is about a fifth of what Tavus does.
- Why he hates the word "avatar." Putting a face on "some shitty LLM audio pipeline" is, in his words, more harmful than helpful. He wants AI humans with memory that build connection over time, not a face on a script.
- Four classes of in-house models. Tavus is vertically integrated because nothing off the shelf was good enough. To get close to the Turing test, perception and understanding have to be tightly coupled to rendering, with shared context and embeddings.
- They generate their own training data. Tavus runs a studio to capture two-way video — what was said, how the other person reacted, the micro-expressions. The data that doesn't exist anywhere (Hassaan's example: people on a first date) they have to source themselves. The half-serious idea that came out of it: a Tavus dating café.
- The pivot that meant firing customers. After the Series A, Tavus had drifted into AI sales tools. Hassaan looked at the team, decided no one cared about sales tooling, and returned to the original vision. He calls the investor conversation one of the scariest he's had: "Hassaan, you better freaking be right."
- An '85 Corvette in the office, and a favorite corporate failure. The C4 Corvette as a vintage computer and underdog story. The strategy failure that lives rent-free in his head: Palm — the best product and OS of its time that still lost on bad business decisions.
- Codex or Claude Code? Claude Code, 100 percent. He re-runs the head-to-head every couple of months on identical prompts; Claude Code keeps winning.
- The use case he didn't see coming. Elderly care companions — an always-present AI that knows someone, sees them, and helps caretakers who are thousands of miles away.
More about Hassaan: LinkedIn, X, Tavus
Chapters
- 0:00Cold open
- 0:14Who Tavus is
- 0:39Where Tavus sits vs Synthesia, HeyGen, Hedra, and Google
- 1:36How customers use AI humans
- 2:03Why he hates the word "avatar"
- 2:38Building four classes of in-house models
- 3:18Generating their own training data
- 5:08The pivot that meant firing the first customers
- 5:59The investor conversation
- 6:10The '85 Corvette in the office
- 7:03Favorite corporate strategy failure: Palm
- 7:50Codex or Claude Code?
- 8:33The use case he didn't see coming
- 9:21What has to become true for Tavus to win