The server
We work out how big it needs to be, help you buy it (or rent it), then install the inference software. Usually vLLM or llama.cpp, with an open model such as Llama, Mistral, Qwen or Gemma.
Private AI
AI that runs on your own hardware
We pick the hardware, install an open model and connect it to your files. Staff log in with their normal work account and ask questions in plain language. Each answer links to the paragraph it came from, and people only get answers from documents they're allowed to open anyway.
Why not just use a cloud AI service?
For a lot of work you can. But law firms, accountants and anyone else holding client files tend to stop at the same point: they don't want those files uploaded to a service they don't control, and they'd like to see who asked what. Open models have got good enough for drafting, summaries, translation and searching your own documents. And they run on one decent GPU.
What's included
What we do
Search over your documents
Contracts, case files, manuals, policies. The system indexes them where they already live and keeps the same permissions as the file share.
A chat screen for staff
Something that looks like the AI tools people already know, behind your normal login, with shared prompts for recurring tasks and a log of what was asked.
Agents for routine work
Sorting incoming mail, pulling figures out of PDFs, drafting a first version of a standard letter. A person checks the result before it goes anywhere.
Connections
File servers, SharePoint, Google Drive, mail and practice-management software, through whatever API they offer.
Rules and training
A short usage policy, how long logs are kept, a test set of your own questions to compare models, and a session for staff (the EU AI Act now expects some AI literacy).
What you have at the end
- The running system, monitored and backed up
- An admin handbook and a diagram of how it fits together
- Test results on your own questions
- A training session
What we don't do
- Training a model from scratch
- Selling you hardware at a margin. You buy it in your name
How it fits together
Nothing goes out of the building
The documents stay where they are and get indexed there. When someone asks a question, the system looks up the passages that person is allowed to read and hands them to the model with the question. The answer comes back with links to those passages.
- Permissions come from the file share or document system
- Questions and answers are logged, so you can check them
- A better model next year can replace this one without starting over
Sizing
How big a machine?
Rough starting points. We size it properly after looking at how you'll use it. What fills the GPU is the size of the model and how many long conversations run at once. The number of accounts matters much less.
| Team | Model size | Hardware |
|---|---|---|
| Up to 10 people | 7–14 billion parameters, quantised | One GPU with 16–24 GB |
| 10 to 50 people | 24–32 billion parameters, quantised | One or two GPUs, 48 GB in total |
| 50 or more, or lots of agents | 70 billion parameters | Two to four GPUs, 80 GB or more |
Prices
What it costs
Pilot
from €5,900
Four to six weeks, one use case, one team. Hardware is extra.
Production setup
from €14,500
Installation, connection to your document stores, permissions and staff training. Hardware is extra.
Running it
€290 / month
Monitoring, backups, security updates and model updates.
Prices are without VAT. Anything else is quoted, or charged at €96.80 an hour. All prices.
Software
What we use for this
- NVIDIA CUDA
- vLLM
- llama.cpp
- Ollama
- Open WebUI
- PostgreSQL + pgvector
- Qdrant
- Keycloak
- Microsoft Entra ID
- Docker
- Ubuntu
Questions
Things people ask
- Does anything leave our network?
- No. The model, your documents, the search index and the logs all stay on your hardware. Updates are downloaded when we decide to, and can be copied in offline.
- Is a local model as good as the big cloud ones?
- For summaries, drafting, translation and questions about your own files, close enough that most people won't notice. For hard reasoning problems the largest cloud models are still ahead. We test both on your own questions before you spend anything.
- Can we try it first?
- Yes, that's how we prefer to start. A pilot is one team and one use case, four to six weeks, on a single GPU workstation or a rented server.
- What does it cost to keep running?
- Electricity, our monthly fee and the occasional hardware upgrade. There's no charge per user or per question.
Articles
If you want the detail
Private AI
How big does a local AI server need to be?
How much GPU memory a private language model needs, why the number of people using it matters less than you'd think, and how big the document index gets. A way of working it out, instead of a shopping list.
Managed servers
What we let AI agents do on a server, and what we don't
Agents are good at watching servers, linking up symptoms and writing the fix. The hard part is deciding what they may do without asking. The four groups we use, and the safeguards around them.
Got something that needs looking after?
Tell us what you have and what's bothering you. You'll hear back within one working day, usually with a few questions before any price.