Reference Architecture
A typical sovereign AI stack, sized to your team and kept entirely inside your network.
Hardware tiers
| Tier | Best for | Typical hardware class | Notes |
|---|---|---|---|
| Small practice | 1–10 users, light document and note workflows | Single high-memory GPU workstation | Quiet, office-friendly; no server closet required. |
| Mid-size | 11–50 users, multiple departments | Dedicated on-prem server or locked rack | Redundant storage, scheduled backups, centralized access. |
| Multi-site | 50+ users, multiple locations | Rack deployment with site-to-site access controls | Segmented by role and location; IT-managed. |
Exact models and specifications are tailored per engagement. We do not resell hardware; we provide procurement guidance your IT team can validate.
Software stack overview
The stack is intentionally generic: open-weight models, local inference, and a retrieval layer that draws from documents you already own. No cloud API calls during normal operation.
- Local inference server: hosts open-weight language models on your GPU/CPU.
- Internal web interface: staff interact through a browser on the local network only.
- Retrieval/knowledge base: connects to your approved internal documents, policies, and templates — not the public internet.
- Audit & access layer: authentication, role-based access, and query logging.
Network isolation
The diagram below shows the typical data path: your staff's prompts and documents never leave your LAN. There is no outbound route to a third-party AI service.
Updates & maintenance
- Models update on your schedule, not silently.
- We document the update procedure and rollback plan before any change.
- Knowledge-base documents are indexed from locations you control; no external sync required.