Why I'm Betting on Real-Time AI
Five months ago I joined a startup building photorealistic AI avatars. People keep asking me why. This is the long answer.
The short version
I believe real-time, embodied AI will be how most people interact with machines within a decade. Not chatbots or voice assistants trapped behind a speaker grille, but human-like presences you talk to the way you'd talk to a colleague.
Anam is building the platform for that. I wanted to be part of it.
What I mean by real-time
The word "real-time" gets thrown around loosely in tech. A dashboard that updates every five seconds is not real-time. A chatbot that responds in two seconds is not real-time, not in the sense that matters for human conversation.
Human conversation has a natural rhythm. We expect responses within 200–400 milliseconds. Go beyond that and the interaction feels sluggish. Get below 200ms and it starts to feel conversational. Get below 100ms and it feels seamless, like talking to another person.
200ms
Conversational threshold
100ms
Seamless threshold
~300ms
Typical human turn-taking gap
The bar is a delay the human brain doesn't register, not "fast enough for a demo". That's an engineering problem of a completely different order to what most AI companies are solving.
Why embodiment matters
There's a common assumption that text and voice are sufficient for AI interaction. For many use cases, they are. But for teaching, coaching, customer service, companionship, visual presence changes everything.
We're social animals: we read faces, mirror expressions, and trust people who hold eye contact. That's deep evolutionary wiring, not a nice-to-have. Strip it away and you lose something real about communication.
I spent time in consulting watching companies deploy chatbots that technically solved the problem but that users avoided. The technology worked. The interaction didn't. What was missing was almost always presence: the feeling that you're talking to something that's actually paying attention to you.
The technical challenge
Building a photorealistic avatar that responds in real-time is genuinely hard. You need to solve several problems at once:
Generation quality. The avatar needs to look real, not "good for AI" real. Anything less triggers the uncanny valley and you've lost the user before you've started.
Lip synchronisation. Mouth movements need to match generated speech with frame-level accuracy. Humans are very good at spotting lip-sync errors; we've been reading lips since infancy.
Expression matching. Facial expressions need to match the content and tone of what's being said. A customer service avatar that smiles while delivering bad news is worse than no avatar at all.
Latency. Everything above needs to happen within that 200ms window. Every millisecond spent on generation is a millisecond not available for inference, network transit, or audio synthesis.
These constraints pull against each other: push quality higher and latency suffers, optimise for speed and quality drops. The hard part is making everything work at once, at scale, in a browser.
Why now
A few things lined up to make this the right moment.
Model efficiency. Generative models have reached the point where photorealistic face generation is possible on consumer hardware. Two years ago this required a data centre. Now it runs in a browser.
The plumbing exists. WebRTC, edge computing, and streaming inference have all matured enough that sub-200ms round trips are achievable at scale.
People actually want this. Every large enterprise I worked with in consulting was actively looking for better ways to deploy AI-powered interactions. The demand is real and immediate.
What I'm actually doing
I joined as Founding Product Engineer. In practice, at a startup of this size, that means I'm involved in everything from technical architecture to customer conversations to hiring.
The thing I'm most focused on right now is the developer experience. We're building a platform, not an application. Our success depends on other developers being able to build compelling experiences on top of our technology. That means great SDKs, clear documentation, and sensible defaults.
The bet
I'm not naive about the risks. Most startups fail. The technical challenges are real. The market is early and unpredictable. There are well-funded competitors.
But I keep noticing the same thing: the companies that win are the ones that solve the interaction problem, not the intelligence problem. Current LLMs are astonishing and still trapped behind a text box. The next step is giving AI a face and a voice, with the speed to respond like a human.
That's what we're building. I think it matters.