The mouth
is doing
the talking.
The eyes are doing
the lying.

Every avatar on the internet
is dying in the same place.

Twenty-two years into the synthetic-face arms race, we have rendered pores. We have rendered subsurface scattering. We have rendered the translucence of a soft earlobe lit from behind. We have spent eleven figures teaching computers to draw skin and hair. And yet every single avatar — every VTuber, every brand spokesperson, every NPC, every AI-generated talking head — fails in the same six square centimetres.

The eye region.

It's not subtle. It's the first thing a human notices and the last thing rigs solve. We notice in twelve milliseconds, faster than conscious thought. Twelve milliseconds. If the eyes are off, the rest of the face doesn't get a second chance.

We can render the texture of a freckle.
We still can't make an avatar look back.

Pretty face. Dead stare.

The current state of the art does one trick: blink every four seconds, rotate the eyeballs toward a script-defined target. Wash. Repeat. We call this "facial animation" because we don't know what else to call it. It is the cargo cult of human-ness. It is what you get when the animator's checklist says "eyes — done" before the eyes have ever been thought about.

Real eyes do not behave that way. Real eyes saccade. Three to four micro-jumps per second, sub-frame, ballistic. Real pupils dilate when something interesting walks into the room and constrict when light hits them. Real eyelids blink asymmetrically. Real gaze converges when you look at something close. Every one of these signals is a tell, and a face that ignores them is a face we read as not-a-face.

A face is a frame.
The frame is the iris.

We built Eyeconic because we got tired of waiting for the avatar industry to fix the one thing that matters. We don't render skin. We don't generate voice. We don't compose scenes. We send you back a single payload — fifty-two ARKit-compatible shapes, per-eye gaze vectors, head pose, blink state, dilation, squint — derived from whatever camera you have, in fifteen milliseconds.

That's the entire product. We are not trying to be a 3D engine. We are not trying to be a voice clone. We are not trying to be a digital double. We are one eye region wired correctly, and the rest of the rig is yours.

Other APIs detect a face.
We detect a glance.

If you sell attention
you are our customer.

VTubers whose audience pays in superchats. Virtual influencers whose face has to ship eleven markets without going cross-eyed in one of them. Game studios who want NPCs that notice the player before the player notices the NPC. Brands who want a spokesperson whose eye-line is locked across every campaign, every render, every language dub. Researchers building social robots that have to pass the third-second test.

You all share the same bottleneck. You all run into it on the same day. You all solve it with the same expensive hack — a VFX team in post, a half-day shoot for "eye plates," a junior animator hand-keying the gaze. We are replacing all of that with one HTTP request.

An iris is a moat.

Models commoditise. Pipelines commoditise. Voice clones commoditise. The thing that doesn't commoditise is perception — the millisecond a human decides they are being looked at, or not. That decision happens before language, before brand, before product. It happens in the eye.

We're building the only API in the world that takes that decision seriously. We think in five years there is exactly one company sitting between "video frame" and "rig payload" for awake avatars. We intend to be it.

⟶ Eyeconic · Frankfurt · 2026