Deploy
Deployment shapes
Hosted SaaS, on-prem in your own cloud account, or an appliance on your LAN — the same product, three delivery models.
The same software ships three ways. The choice is driven by where your data is allowed to live and where your users physically are.
| Hosted (SaaS) | On-prem cloud | Kiosk / LAN appliance | |
|---|---|---|---|
| Runs in | Our infrastructure | Your own cloud project | Your building |
| Who operates it | Us | Us, inside your account | You, with our support |
| Conversation data leaves your control | Yes, to us | No | No |
| Time to first conversation | Days | 2–3 weeks | 3–4 weeks including hardware |
| Best for | Web and app deployments, pilots | Regulated data, public sector | Malls, airports, branches, hospitals |
Hosted SaaS
You call our API; we run the GPUs. Fastest path to a working deployment and the usual starting point for a pilot. Capacity scales with demand automatically.
On-prem cloud
A dedicated stack inside your cloud project — one customer per stack, no shared compute and no shared session state. The standard layout is two virtual machines:
- a proxy host holding the public entry point, TLS, the web front end, and media relay;
- a GPU host running speech, the model, and rendering, with no public IP.
We provision it with Terraform and Ansible, and we operate it with you. Conversation content stays inside your project; our central infrastructure sees licensing and health signals only. See Security and data.
Kiosk / LAN appliance
For physical installations. Two variants:
- Single kiosk — one box behind the screen runs everything. Simplest, and the right answer for one or two screens per site.
- Appliance plus edges — one well-provisioned GPU host on the LAN serves 4–12 kiosk screens. Cheaper per screen, one host to update, and the entire conversation path stays on your local network.
Both variants keep working when the internet link degrades: session state lives on the local network, and the deployment has a defined offline fallback for the screen. Fully air-gapped operation (no periodic licence check at all) is not supported yet — talk to us if that is a hard requirement, as it changes how model weights are delivered.
Moving between shapes
Pilots that start hosted routinely graduate to on-prem or kiosk. Avatars, the knowledge base, conversation flows, and integrations carry over — the migration is a redeploy, not a rebuild.