livepresence.
Private beta Enter the console

Live neural video presence

You speak.
A face answers.

A photoreal persona that hears you, thinks with your model and answers out loud — lip-synced at 25 fps, first frame in about three-quarters of a second, on a single GPU in your rack.

Self-hosted Disclosed on every frame Consent-attested identities
Enter the console Request access

The console is password-gated while we run a private beta.

Try the grounded demo

Nova's already answering as a sample business — ask her anything, or switch businesses anytime. Client-side retrieval; no live model in this page.

Choose a business

Who should Nova answer as?

Pick one and Nova locks to that business's knowledge base for a three-minute session.

Grounded demo · client-side. Retrieved from each business's demo knowledge base, in your browser. In production: your model, your documents.

the console, annotateda – e · how the live console works

Hands-free mic

tap once. voice activity detection segments your speech; it answers whenever you pause.

Barge-in

interrupt mid-sentence. the render yields and the mic is already listening.

Disclosed, always

the badge is in the ui. the watermark is in the pixels of every rendered frame.

Live caption

whisper large-v3, self-hosted, writes both sides down beside the call.

Telemetry

last-turn latency measured and shown. this machine keeps its own receipts.

measured / first frameproduction evals — not a bench demo

750

msp50 · first frame

median time from the first word of the spoken answer to the first rendered video frame. inside the render worker alone: 250–400 ms. under a second, the face is simply there.

25 fps

photoreal neural render, lip-synced live

large-v3

whisper asr · self-hosted transcription

1 gpu

llm + renderer + asr on a single card

250–400 ms

render-worker first frame, measured

The call

A conversation, not a form.

No push-to-talk. No send button. You talk to it the way you talk to a person on a video call — including cutting it off.

Listening

Hands-free after one tap

Tap the mic once and the line stays open. Voice activity detection finds the edges of your sentences, and the persona answers whenever you pause. Tap again to stop.

Barge in mid-sentence

Start talking over the answer and the render yields immediately — the turn is abandoned, the mic is yours. Interruptions are part of the protocol, not an error state.

Every word on the record

Whisper large-v3 runs beside the call — self-hosted, like everything else. A live transcript tracks both sides, and a text composer is there when you'd rather type.

One turn

What happens when you pause.

The one genuinely sequential thing on this page. Four stages, one card, under a second to first frame.

01VAD

Speak

The mic is already open. Voice activity detection segments your speech and closes the turn the moment you stop.

02whisper large-v3

Transcribe

Whisper large-v3, on your card. Your words land in the live transcript as the turn closes.

03your endpoint

Think

Any OpenAI-compatible endpoint — bring your own brain. Production runs a 30B model on vLLM, on the same GPU.

04250–400 ms

Render

The neural renderer starts speaking the reply — first frame in 250–400 ms, lip-synced at 25 fps, watermarked.

Barge-in, at any stage. Talk over the answer and the pipeline drops the turn mid-render and goes back to listening.

Identities

A presence from a single portrait.

Nova and superNova ship with the beta. Both are fully synthetic — there is no real person behind either face.

Nova — a fully synthetic persona portrait: a woman with auburn hair and a navy sweater, emerging from the dark page ground

Nova.

AI GENERATED

The shipped default. Calm register, conversational pace — the face most beta sessions start with.

sha-256 bc585942…66ab1dc4 · fingerprint of this portrait

superNova — a fully synthetic persona portrait: a woman with dark hair in a bronze silk top, emerging from the dark page ground

superNova.

AI GENERATED

The second shipped identity — same pipeline, different presence. Generated, attested, fingerprinted.

sha-256 8eede8a7…d440dc4c · fingerprint of this portrait

Yours takes one photograph and one promise.

A single front-facing portrait is enough — the engine builds a persona that speaks, listens and holds your gaze. But nothing renders until the paperwork is real:

  • One portrait — front-facing, even light. JPEG, PNG or WebP. That's the entire input.
  • An explicit consent attestation — you confirm you have permission to use this person's likeness. No attestation, no persona.
  • A SHA-256 fingerprint of the source image, recorded at upload — every identity traces back to exactly one attested file.

Trust & provenance

Disclosed by design.

Synthetic media you can put in front of customers — because every frame says what it is, and every identity can prove where it came from.

WATERMARK

Every frame carries the disclosure

Rendered output ships with an AI GENERATED watermark burned into the pixels, and the console wears the same badge. Transparency is a feature here, not fine print.

always on
PROVENANCE

Signed at the source

Every source image gets an HMAC-signed provenance record and a SHA-256 fingerprint. If a face renders, you can show exactly which portrait it came from and who attested to it.

always on
CONSENT

Likeness requires permission

No persona is created without an explicit consent attestation, recorded with the fingerprint. Fully synthetic personas — faces that belong to no one — are a first-class alternative.

enforced at upload
FAIL-CLOSED

Age assurance that defaults to no

An age-assurance classifier gates every adult asset: no verdict means no asset. The 18+ tier is synthetic-only, entitlement-gated and off by default, with youth-term filters throughout.

fail-closed

// the safest failure mode is the one that was designed first.

Protected runtime

Nothing leaves the box.

The whole stack — language model, neural renderer, speech recognition — runs on one RTX-class GPU in your rack. Your voice is transcribed on your card. The reply is generated on your card. The face is rendered on your card.

There is no third-party inference API in the loop and no transcript mirrored to someone else's cloud. Point the engine at any OpenAI-compatible endpoint — ours runs a 30B model on vLLM, sharing the same silicon as the renderer.

Commercial API

Script in, watermarked video out.

One endpoint. Send text and an avatar, get back a finished, watermarked render — the same engine that runs the live call, driven by your systems. The request below is the contract as it ships, field for field.

Billable seconds are the rendered video seconds — nothing else. No per-request fees, no rounding games.
●  requestPOST → 202 accepted
POST /api/v1/generations
Authorization: Bearer lp_live_…
Idempotency-Key: 2026-08-16-supernova-001
Content-Type: application/json

{
  "avatar_id":      "b9ee8505-d424-433d-b53e-49d913e45b51",
  "text":           "Welcome back. Good to see you again.",
  "input_mode":     "script",
  "render_profile": "realtime_hq",
  "content_rating": "general"
}
●  poll GET /api/v1/generations/{id}completed
{
  "id":               "gen_…",
  "status":           "completed",
  "avatar_id":        "b9ee8505-…",
  "render_profile":   "realtime_hq",
  "content_rating":   "general",
  "output_text":      "Welcome back. Good to see you again.",
  "video_url":        "/api/v1/generations/gen_…/video",
  "billable_seconds": 1.6,  // == rendered video seconds
  "first_frame_ms":   742,
  "total_ms":         1840
}

Private beta

The beta is private. The machine is real.

The console is running right now, behind a password. Tell us what you would point it at — a human reads every request.

needs an @ and a dotted domain — e.g. [email protected]
posts to /v1/demo-request · falls back to plain email
Couldn't reach the machine from here — this page may be running as a static preview. Your note is ready to go the old way instead: send it by email →
Received. Your request is on the machine — a human replies from [email protected], usually within a day.

Already holding a seat? Enter the console →