Tavus

As AI gets better at simulating human conversation, the question that everyone asks isn't around what the technology can do, it's what happens to the humans on the other side of it. That's the conversation we're bringing to HumanX week alongside Orum 🥇 and Rox in SF. Orum uses Tavus to power AI roleplay personas inside its sales coaching suite where reps practice real conversations with AI that can see them, hear their tone, and react with genuine emotion before they ever pick up the phone. Rox is extending that vision by deploying autonomous agents across the entire revenue cycle, and together we're exploring where the human element becomes more important as AI takes on more of the workflow, not less. This matters because the future of AI in learning and development isn't about replacing the human, it's about making the practice so real that when the human shows up for the actual conversation, they're ready. Eighty open spots, four speakers seated among guests, drinks and food available, just a few blocks from the HumanX conference. We would love to see you there next Wednesday at 6:30pm if you're in the area. Invites are limited so make sure to request yours here asap: https://luma.com/f6etclpo

We're thrilled to announce that OpenClaw now integrates with OpenPAL, our newest product release. OpenPAL is your 24/7 real-time AI assistant that handles every task you hate doing yourself, and a few you didn't realize you were avoiding. With our partner integration with Apple, just ask Siri to install it directly onto your device. [Voice prompt: "Hey Siri, do daa thang."] With our integration with Theranos, you'll never have to go to the doctor again because your OpenPAL will do it for you. And with our newest plug into Ramp, OpenPAL can become your next CFO in seconds. If you're looking for full time freedom in your life, sign up for OpenPAL today.

You can tell when someone is losing interest in a conversation before they say a single word about it because you can see it in their face, hear it in their tone, and feel the shift in energy. That's what was truly missing from AI until our recent model release, Phoenix-4. Previous models, including our own Phoenix-3, could move a face according to audio but couldn't actually feel the conversation they were in. Phoenix-4 is a completely new architecture that generates and controls real emotional states in real-time, listens and reacts while the user is still speaking, and reflects genuine understanding back through every micro-expression and response. This matters because the moment an AI video agent can show concern when a patient describes pain rather than smiling through it, that patient is more likely to open up, follow their treatment plan, and come back. When a sales agent reads hesitation in a prospect's voice and adapts before the objection is even spoken, conversion rates climb and conversations last longer. When a coaching agent mirrors the emotional weight of what's being practiced rather than staying neutral, learners retain more and come back to practice again. Companies building with face-to-face AI are already seeing up to 3x faster sales ramp, 54% increases in learner confidence, and 65% user retention after 30 days. We built Phoenix-4 for that exact moment when someone stops thinking about the technology and starts trusting the conversation. Try the demo for yourself here: https://lnkd.in/gWpvcvB9

When someone says "yeah, I'm fine" while their voice drops and they look away from the screen, most AI takes them at their word because all it has is the transcript. The tone, the expression, the disconnect between what was said and what was meant is invisible to it, and that's where every important moment in a conversation lives. This is the understanding gap in conversational AI today. Systems convert speech into text and throw away roughly 70% of the signal that actually carries meaning. Sarcasm becomes indistinguishable from sincerity, hesitation gets read as confidence, and frustration paired with hope in the same breath gets flattened into a single label that captures neither. Raven-1 is our multimodal perception model built to close that gap. Instead of analyzing audio and visual signals separately, it fuses them together in real time to produce rich natural language descriptions of what the person is actually feeling, descriptions that your LLM can reason over directly. Sub-100ms audio perception with context that's never more than 300ms stale. This is also why video matters so much more than voice alone for high-stakes conversations. Voice gives you tone, but it can't show you someone looking away, leaning back, or visibly struggling with what they're about to say. Raven-1 sees all of it, and that's the difference between AI that listens and AI that actually understands. Swipe through below to see how it works.

Our CEO Hassaan Raza asked Joshua Bailey to interview the newest member of our sales team, and let's just say Joshua may have overestimated how much the interview was about him. What he missed is something a lot of companies miss when they think about AI in sales: the value isn't in talking more or information dumping on a prospect, it's in understanding what the person on the other side of the conversation is actually feeling and adapting to it in real time. An interactive AI SDR that sits on your company page that can pick up on a website visitors hesitation, sense when interest is building or fading, and shift its approach based on emotional context rather than a rigid script turns what would normally feels like a pitch slap into a conversation someone actually wants to continue. That's what a full-stack AI video agent from Tavus can bring to your website or app. Perception that reads tone, expression, and intent together. Personality that shapes how the AI Human shows up for each unique visitor. And memory that makes every follow-up feel like a continuation, not a cold start. Prospects don't convert because they were talked at. They convert because they felt understood. Joshua is still working on that part.

Emotional intelligence is quickly becoming the most important layer in conversational AI, and most systems are still operating without it entirely. Today's AI can generate a perfectly worded response to someone who just told them they lost their job, and deliver it with a smile because the system has no idea what the person on the other side is actually feeling. It processes words, not meaning. It hears language, not the hesitation in someone's voice or the way their expression shifts mid-sentence when a topic gets heavy. Our Raven-1 model is how we solve this at Tavus. It fuses audio and visual signals together in real-time so our AI video agents aren't guessing at emotional context, it's reading tone, facial expression, and intent as a single continuous signal the same way you would if you were sitting across from someone. When that perception layer feeds into how the agent responds, remembers, and adapts its personality over time, the entire conversation changes. This matters because the use cases that need AI to build trust, whether that's healthcare, simulation training, sales, or coaching, depend on something deeper than a good answer. They depend on the person feeling understood before the AI even responds. Try the Raven-1 demo for yourself: https://lnkd.in/dBT8435i

A few weeks back we hosted another dinner in SF and the people in the room reminded us exactly why we keep doing these. We're building technology that teaches machines to deeply understand us, remember us, and build real relationships with us over time, not just look and sound like us but actually know who they're talking to. The emotional intelligence to read how someone is feeling before they say it, the personality and memory that make a conversation feel like it matters, and the judgment to know when to push back and when to just listen. The only way we know how to get that right is by staying close to what real human connection actually looks like. That's what happens around a dinner table with people who share the same passion for building in the real-time AI space. Someone shares an idea they've been sitting on for weeks and the energy in the room shifts. Someone pushes back on an assumption and the conversation gets deeper because of it. Someone remembers what you said last time and picks the thread back up. These are the dynamics we're teaching machines to replicate, and we learn something new about them every time we sit across from each other. If you're building in this space or just thinking about where human-AI interaction is headed, we'd love to have you at the next one.

Teaching machines the art of being human requires more than a face. Raven-1 handles perception, fusing audio and visual signals in real time to understand not just what someone says but how they say it, how they look when they say it, and what that combination actually means. Phoenix-4 generates real-time facial behavior with genuine emotion, reflecting all of that understanding back with presence you can feel. Sparrow-1 handles conversational timing, knowing when to respond, when to wait, and how to recover from an interruption the way a real person would. But models alone aren't enough. What makes the experience of talking to an AI video agent feel human is what sits between them: the personality that shapes how the agent responds to you specifically, the memory that remembers what you talked about last week, and the judgment that ties it all together so the conversation builds trust over time. Perception. Intelligence. Personality. Rendering. Every layer, working together. Try each model for yourself, links in the comments.

We're building every layer of the AI Human, not just the face. At Tavus, we spent years developing the models that give machines the ability to see, hear, and respond with real presence, Phoenix-4 for real-time rendering and emotion, Raven-1 for perception, Sparrow-1 for conversational timing. But what we've learned is that a well-rendered face on top of a robot is still a robot, which is why we now build the layers underneath it too: deep personality systems, memory that carries across conversations, judgment that knows when to push back and when to listen, and intelligence that ties all of your context together so the machine isn't just reacting but actually understanding. That vision is what our co-founders Hassaan Raza and Quinn Favret have been pushing since day one, and it's showing up everywhere. Hassaan joined SAP on stage at their brand new Experience Center in Palo Alto where every customer who walks through the doors now interacts with our AI Humans firsthand. Quinn brought a real-time AI replica of B.B. King to the stage at Entech LA to show what's possible when perception, personality, and rendering all work together. And right now, our AI Humans are being showcased at GTC in the Amazon Web Services (AWS) booth, showing people firsthand how interactive video agents work in production. If we're going to have real-time video agents that are assistants and coworkers that people actually trust, they have to feel like there's someone on the other side who knows them. That requires building all the layers, not just the visible ones.