Tavus.

Here is a thing that is true about computers: for roughly seventy years, using one has meant translating what you want into a form the machine can accept. You type. You click. You learn the syntax. The machine, for its part, has never once looked up from the desk to see whether you were confused. This is a strange arrangement if you think about it, which mostly nobody does, because it is simply how things are.

Tavus, a San Francisco company founded in 2021 by Hassaan Raza and Quinn Favret, thinks this arrangement is about to end. Its pitch, which it calls "human computing," is that using a computer should feel as natural as talking to a friend or a coworker - meaning you talk, it listens, it watches your face, and it answers in a voice with a face attached, in real time, with the timing of an actual conversation. If that sounds like an ambitious way to describe a video call, that is sort of the point. The other party on the call is not a person.

The mechanism is a product Tavus calls the Conversational Video Interface, or CVI. It is an end-to-end pipeline for face-to-face AI: it handles perception (watching and hearing you), dialogue (deciding what to say), and rendering (producing a lifelike talking face), and it does the whole loop in under a second. Under a second matters more than it sounds. Above roughly one second of delay, a conversation stops feeling like a conversation and starts feeling like you are waiting on a machine. Below it, something flips, and your brain grants the thing on screen the benefit of the doubt. Tavus has organized an entire company around staying on the right side of that threshold.

It is worth being precise about what Tavus is and is not. It is not, primarily, a service that makes a talking-head video of your CEO for a corporate all-hands, though it started closer to that. In its earlier life Tavus did generative AI video - face and voice cloning to produce personalized videos at scale, the kind of thing where a sales rep records one clip and the system generates a thousand versions with each prospect's name in them. That was the wedge. The company raised an $18 million Series A in March 2024, led by Scale Venture Partners, and used it to open the platform to developers through an API.

What Tavus discovered, as companies with a good wedge often do, is that the more interesting product was the live version. Not a video you watch, but a conversation you have. So the destination became CVI, and the thing you talk to became a "PAL" - Personified Application Layer, which is a backronym doing a lot of work, but the idea underneath it is coherent. A PAL is an application you converse with face-to-face. It maintains a lifelike visual presence, reads your expressions and gestures, understands emotion and timing, remembers context, and moves between video, voice, and text as needed. It can also do things: book a meeting, pull a record, submit a form, call an external API mid-sentence, because the underlying model decides when to reach for a tool.

This is the part that separates Tavus from a novelty. An avatar reads a script. A PAL reads the room. The difference is the entire business.

There are three of them, and, in a naming decision that is either charming or a sign of engineers left unsupervised, they are all birds. Phoenix-4 is the rendering model - it drives lifelike expression, head-pose control, and emotion generation at conversational latency, which is to say it produces the face and makes the face feel present rather than uncanny. Raven-1 is the perception model - it interprets context, people, environments, emotions, expressions, and gestures, reading your gaze and the objects around you in real time. Sparrow-1 is the audio understanding model - it manages timing, tone, intent, and turn-taking, so the PAL knows when you have finished a thought and when you are just pausing. Perception, dialogue, rendering. Raven watches, Sparrow listens, Phoenix speaks.

Building your own models is a choice with consequences. It is more expensive and slower than wiring together somebody else's, and it is the reason a company selling what looks like a friendly video widget describes itself as a research lab. But it is also the moat. If your whole pitch is that the experience feels human - that it lands on the right side of the one-second line and doesn't tip into the uncanny valley - then you cannot outsource the parts that determine whether it does. Tavus lets developers swap in their own large language model, their own voice, their own knowledge base; it does not let them swap in the perception and rendering, because that is the company.

Some parts it does partner on, sensibly. Cartesia supplies fast voice generation, and Tavus has marketed the result as the world's fastest conversational video interface powered by that pairing. Daily, the real-time video infrastructure company, provides the transport layer developers can build on when they want full control of the interface. The pattern is consistent: Tavus owns perception and rendering - the parts that decide whether the thing feels alive - and buys or plugs in the commodity plumbing around them. That is the correct instinct. You want to spend your scarce engineering hours on the difference that customers can feel, not on re-implementing a video pipe that already exists.

On the ergonomics side, Tavus has done the work to be genuinely easy to adopt, which is not always true of research-led companies. There is a CVI React component library with prebuilt, themeable components and hooks for people who live in front-end code. There is an iframe embed for static sites and quick demos, where you paste a snippet and a talking agent appears. There is vanilla JavaScript, a Node and Express path for dynamic embedding, and the Daily SDK route for teams that want to rebuild the whole interface. You can upload PDFs and documents or point the system at a website to crawl, and it retrieves against that knowledge in around thirty milliseconds. The spread of on-ramps - from "drop in an iframe" to "wire up the transport yourself" - is a tell that Tavus has thought about the full range of who shows up, from a solo builder on a weekend to an enterprise team with a compliance department.

The customers are developers and the enterprises they work for, and the use cases are the ones you would guess once you accept the premise. Scalable one-to-one sales coaching. Customer support that looks you in the eye instead of making you read a chat transcript. Candidate screening at volume without a human recruiter on every call. Education, corporate training, healthcare intake, entertainment. The through-line is the same primitive - a live, face-to-face AI agent - pointed at different problems. This is the good kind of software business, where you build one hard thing and sell it many times, and Tavus has been disciplined about keeping the primitive singular and the applications plural.

In November 2025 the market signed off on the thesis. Tavus raised a $40 million Series B led by CRV, with Scale Venture Partners, Sequoia Capital, Y Combinator, HubSpot Ventures, and Flex Capital participating. That brings total funding to roughly $82 million. The framing on the round was deliberately grand - "the next frontier of intelligence: human computing" - and you can roll your eyes at the phrasing while still noticing that the investors backing it are not naive about AI. Sequoia and Scale have seen enough demos to be hard to impress.

There is a version of this story that is uncomfortable, and Tavus, to its credit, does not pretend otherwise. An AI that reads your face and responds with calibrated emotion is a powerful thing to hand to a sales funnel. Consent, likeness, and trust get more complicated when presence becomes an API call - when a face that looks you in the eye can be spun up on demand and pointed at whatever the customer wants. These are real questions, and they do not have tidy answers yet. What Tavus has done is make the underlying capability good enough that the questions are now worth asking, which is a strange kind of achievement but a genuine one.

The honest way to describe Tavus is this: it is a well-funded, research-heavy company making a specific and falsifiable bet. The bet is that the interface for the AI era will not be a chat box but a face, and that the company which controls the perception-and-rendering layer of that face will own something valuable. Maybe that is right and maybe the chat box wins because typing is quietly what a lot of people prefer. But if you are going to bet on the face, you want to be the one that got the latency under a second and built the birds. Tavus is, so far, that one.