Your own AI, inside your own walls.
A secure AI server sized to your company — one machine or a rack — running models trained on your material, with nothing leaving your network.
Most AI adoption means posting your work to somebody else’s computer. For any business holding client confidences, unreleased material, personal data, legal files or anything under an NDA, that is the single thing standing between them and the whole technology.
A private AI server removes the question. The model runs on hardware you own, on a network you control. Prompts, documents and outputs stay inside. It can sit behind your firewall, on an isolated VLAN, or fully air-gapped with no route out at all.
The part that makes it worth owning is the training. A general model knows the world; it does not know your pricing, your contracts, your clients, your past projects or how your company phrases things. We build each one against the business it serves, so it answers from your material rather than from the internet’s average.
What we build
Hardware sized to the company
From a single GPU workstation for a small team to a multi-GPU rack serving a whole organisation. Specified against how many people use it at once and how large a model the work genuinely needs — not against the biggest number in the brochure.
The model stack, self-hosted
Open-weight models served locally with a proper inference layer, an OpenAI-compatible endpoint so existing tools can point at it, and a chat interface your staff can actually use. Models are swappable as better ones land.
Trained on your business
Your documents, contracts, catalogues, transcripts and history indexed into retrieval, with fine-tuning on top where tone and format matter. Answers come with citations back to the source document, so staff can check rather than trust.
Access, audit and control
Single sign-on, per-department permissions so finance material is not answerable to everyone, and a full audit log of who asked what. The things that make it approvable by whoever signs off on risk.
How we work
Start from the work, not the hardware
We map what people will actually use it for before specifying anything — which routinely lands a sharper system on smaller, cheaper hardware than anyone walked in expecting.
Prove it on your data first
A pilot on a slice of your real material, measured against the questions your staff actually ask, so the specification follows evidence rather than a brochure.
Yours to keep running
Documented, monitored and handed over with the runbook. We can maintain it and keep the models current, but you own the hardware, the weights and the index, and nothing phones home.
Sizing, roughly
What each tier tends to suit. Indicative only — the specification follows the pilot, not this table.
| Tier | Concurrent users | Suits | Typical use |
|---|---|---|---|
| Single workstation | 1-10 | Small firms, one department, a practice | Document Q&A, drafting, summarising, internal search |
| Small server | 10-50 | Mid-sized companies, a few departments | The above plus transcription, extraction and light automation |
| Multi-GPU rack | 50+ | Large organisations, regulated sectors, air-gapped sites | Company-wide assistant, heavy document processing, custom agents |
Concurrency depends on model size and response length. We measure it on your pilot rather than quoting a theoretical figure.
What you get
- Specified and commissioned hardware, in your ownership
- Self-hosted model stack with an OpenAI-compatible endpoint
- Retrieval index built from your own material, with citations
- Fine-tuning where tone, format or domain language requires it
- SSO, role-based permissions and audit logging
- Monitoring, update path and operator documentation
Before you ask
- How does it compare to the big commercial models?
- On the work businesses actually run — answering from your own documents, drafting in your format, extraction, summarising, search across everything you hold — a private model trained on your material beats a general one that has never seen it. It knows your pricing, your contracts and your history, and it answers without any of it leaving the building. For open-ended reasoning at the frontier we can give the system a sanctioned, logged route out to a larger model, so you get both.
- Does it ever send anything out?
- No. The default build has no outbound route for prompts or documents at all, and it can be air-gapped entirely. Any route out is something you switch on deliberately, with its own logging — never a quiet default.
- What happens when a better model comes out?
- You swap it. The retrieval index, permissions and interface are independent of the weights, so a new model is a change of engine rather than a rebuild. That is a large part of why it is built this way.
- Who keeps it running?
- You can, and it is documented so you genuinely are able to. Most clients take a support arrangement for updates, model changes and monitoring — because they want it, not because we built a dependency in.
The rest of the stack
Tell us what you're building.
Send the brief — or the problem, if the brief doesn't exist yet. We come back within one business day with an honest view of how we can help.
Start a project