Skip to content
Self-hosted

The same pipeline, running on your hardware

Every cloud avatar platform requires you to upload your decks, your people and your voice assets to someone else's infrastructure. For most training libraries that is the single reason the project stalls in security review. Self-hosting removes it.

GPU

1× NVIDIA, ≥ 8 GB VRAM

Host

Linux x86_64, Docker + NVIDIA toolkit

Storage

≥ 100 GB for models and intermediates

Network

No inbound access required

Data flow

Where your data actually goes

In the default self-hosted configuration there are exactly two optional egress points — the language model used for script drafting, and the speech model used for narration. You choose which ones those are, and both can be pointed at models you host yourself. Everything else, including the presenter footage and all rendering, stays on your hardware.

Stage Leaves your network
Deck upload No
Slide rendering & text extraction No
Script generation Optional
Text-to-speech No
Lip-sync rendering No
Final encode & storage No

Script drafting can be pointed at a model you already self-host, in which case no stage of the pipeline leaves your network at all.

Fit

Who self-hosts

Usually the trigger is a security questionnaire rather than a preference. If any of these describe your library, the cloud-only tools are likely to fail review.

  • Schools and universities handling student work
  • Corporate L&D with confidential internal material
  • Public-sector teams under data-residency rules
  • Healthcare and finance training libraries
  • Any org whose security review blocks cloud avatar tools
Economics

Why the cost curve is different

Cloud avatar platforms bill per minute of rendered output, because every render costs them GPU time. On your own hardware that cost is electricity and amortisation. The more you convert, the wider the gap — which is exactly why conversion of large existing libraries is the use case where self-hosting wins.

Rule of thumb: if you have more than roughly a hundred decks to convert, or you expect to re-render them as content changes, the hardware pays for itself.

Private beta

Ask for the architecture pack

We will send the deployment topology, the port and volume map, hardware sizing for your library, and what the onboarding session covers.

No spam. One email when the private beta opens.

No credit card. No self-serve signup yet — we onboard pilot teams by hand.