A box in your server room, about the size of a shoebox, running language models locally. You ask questions about your own documents — which contract has a 30-day notice period, what was agreed with a supplier in March — and the answer is built from your files. Nothing is sent to OpenAI, Google or anyone else.
Who it is for
Contract-heavy work
Law firms, tax advisors, clinics, engineering offices, public contractors.
Under a duty of confidentiality
Professional secrecy makes public AI tools a liability rather than a convenience.
Specialist knowledge in the house
The answer sits in a PDF from 2019 that nobody can find any more.
What you get
- An appliance with 24–32 GB of graphics memory, sized to your document volume
- Local models (Llama / Qwen / Mistral class) — no external API call, no account
- A document index over contracts, invoices, scans and mail, with OCR for paper
- A browser interface for staff, integrated with your file and mail platform
- A training session and a written AI policy, so usage is permitted instead of tolerated
- Monitoring, model updates and maintenance
[cn_tco]
How it works
Pilot
We index a sample of your documents and you test it with real questions.
Appliance
The hardware is installed in your building and connected to your network.
Index
Documents and mail are indexed. You decide which folders stay out.
Training
Your staff learns what to ask — and what the system cannot do.
Operate
Updates, monitoring and a quarterly usage report.
Questions
Not to answer questions. The appliance needs a connection only for updates, and you can approve those manually. Every answer names the document it came from and you open the source in one click. It is a search tool that phrases answers, not an authority. We state plainly where it is reliable and where it is not. Data stays inside your building and under your control, which is the simplest possible answer to that question. We also write down the usage rules — an AI policy — before anyone starts working with it.Does it need an internet connection?
How accurate are the answers?
Is this allowed under data protection rules?
Test it on your own documents
We can run a pilot on a sample before you buy any hardware.