Device-rendered avatars
2.5D or 3D, drawn with WebGL. Works with any voice, with no per-minute fees.
Talking AI avatars for websites & apps
A face for your assistant.
Lip-synced to what it actually says.
A face for your voice agent, lip-synced to its voice. Drawn on the device, or photoreal video where realism is worth the cost.
Ultrabot, built by MLAIA Data ScienceLed by Dr. Yochai Edlitz · Ph.D., Weizmann Institute
Two ways to build a face
Render it on the device,
or stream it as video.
The main choice: draw the face on the device, or stream it as video from the cloud.
Photoreal video avatars
Simli, HeyGen or Tavus. More lifelike, but more delay, a per-minute cost and a third party in the loop.
Characters, not only people
Cartoon, animal or robot faces suit children and playful products.
Your own character
A face for your brand, or one made from a photo, each with a matching voice.
What we have built
An avatar assistant
that runs on the phone.
For a healthcare app, currently in development: faces animated by a shader on the device, with the mouth driven by the real speech audio. Users pick a character and the voice follows. The avatar runs offline in the app and makes no network requests. See also voice agents.
We describe client projects in general terms. Details are shared in conversation, subject to confidentiality.
What decides the design
Realism is one factor.
Delay, cost and trust are the others.
A beautiful face that answers late feels worse than a simple one that answers at once.
Delay
Device faces add almost none. Video adds streaming time, measured on real networks.
Cost per conversation
Video is billed per minute; device faces cost nothing extra.
Privacy
A device face sends no audio to a rendering service.
Accessibility
Captions, a stop button, reduced motion, and a no-face option.
Avatar stack optionsthree.js & WebGLAudio-driven visemesReady Player Me & GLB headsReact Native WebViewHeyGenTavusSimliWebRTC
A practical delivery path
Prototype the face early.
Test it with real users.
Audience & tone
Who it talks to, and how realistic it should look.
Prototype
A working face and voice on your target devices.
Connect the agent
Voice agent, captions, controls and fallbacks.
Test & refine
Sessions with real users, then refinements.
Working for your organisation
Your data, your systems,
your customers' languages.
Built for your organisation, it uses your information and tools.
Answers from your knowledge
Answers come from your documents, with access rules per user.
Connected to your systems
Check an order, book an appointment or update the CRM, confirmed by the user and logged.
Reaching your customers
Reminders and follow-ups, only to people who agreed to be contacted.
Hebrew, Arabic, Russian, English
It replies in the user's language, and lip sync works in all of them.
Three ways to build it
Scripted, free-talking,
or a mix of both.
Not every conversation should be open-ended.
Flow-based
Fixed steps and approved wording. Predictable and easy to audit.
AI agent
Open questions, answered from your knowledge.
Hybrid
A fixed flow, with the AI answering questions along the way.
Before we start
Good questions. Straight answers.
Do we need a video avatar service such as HeyGen?
Not necessarily. A device-rendered avatar costs nothing per minute and keeps data off third-party servers. A photoreal video avatar makes sense when realism is central, for example a branded presenter.
Does lip sync work in Hebrew?
Lip sync driven by the audio works for any language the voice speaks. Accuracy is best on languages close to the training data, so we review it on your languages and adjust where needed.
Can the avatar run on mobile phones?
Yes. Our device-rendered avatars run inside a mobile WebView as well as in desktop browsers. We test frame rate and battery on mid-range phones, not only flagships.
Can we use our own character or mascot?
Yes. A 2.5D avatar can be built from a single well-prepared image of the character; a 3D head needs a model with mouth shapes, which we can specify or prepare.
Talk to a person
Tell us what
you’re working on.
Share the problem, the data you have and what success would look like. We’ll discuss a practical next step.
Please don’t include confidential datasets, credentials or patient information.