I am Sara
A photoreal AI avatar that talks to you live — and emails a personalised video within minutes.
On iamsara.io you click Start, grant mic access and talk face-to-face with Sara. She scans your website during the conversation, recognises returning visitors and delivers a personalised video to your inbox afterwards — fully automated.
±250 ms
To the first spoken token
12
Languages in the CMS
20
Architecture Decision Records
±2 min
Personalised video in your inbox
What it does
Talk, don't type
WebRTC stream with lip sync via HeyGen LiveAvatar. Just interrupt Sara — barge-in stops her mid-sentence and she keeps listening.
She speaks before the model finishes
The text stream is cut on sentence boundaries and the first chunk already leaves at a comma — which makes the conversation feel humanly fast.
Live website scan
Mention your site and Sara reads along during the conversation: scraping plus summarisation are injected as a knowledge block into her next turn. She responds with a real observation, not 'is that right?'.
Personalised video funnel
Scrape the site, screenshot it, write the script with Claude, render an avatar over your homepage and email it — a state machine handles it end to end.
Memory across sessions
Returning visitors are recognised and greeted personally, backed by an AI-distilled profile in Postgres.
Manageable without a developer
Block-based CMS with 30+ section types, a knowledge base with automatic gap detection, a persona editor and translation into twelve languages — the operator steers everything.
How we built it
In a spoken conversation every millisecond counts: the speech turns run on a fast Llama model (±250 ms to first token), while Claude does the heavy thinking — video scripts, memory distillation and translations.
The chat route runs on the edge in three regions and streams plain text the avatar speaks immediately; all database writes happen after the answer, so latency stays untouched.
Leads are not captured via function calling but via invisible markers inside the spoken text — model-agnostic and more reliable. A transform stream extracts them before Sara says them out loud and lights up the checkmarks in the UI.
Twenty ADRs document every architecture decision; crons watch leads and sessions, and the database is backed up automatically every week.
Why we built it
An avatar with a face is exposed the moment it feels slow or generic. This build had to prove a spoken AI conversation can feel sub-second — and that the conversation itself can carry the conversion.
The result is a demo that sells itself: visitors experience what an AI avatar could do for their company, and receive the proof by email — about their own website.
Need something like this in your company?