← services

Private AI.
Client data stays in the building.

For law firms, accountants and anyone else who holds confidential client files: an AI model that runs on your own hardware and answers from your own documents.

No account with an AI company, nothing uploaded, nothing sent out. It can run with the internet switched off.

Start a project Book a 15-min intro call

your office · your machine

  1. Your documentsfiles, email, client notes
  2. The modelruns on your hardware
  3. The answerwith the document it came from

Outside AI servicesnot connected

The whole system sits inside your network. There is no line to an outside AI service to switch off, because there isn't one.

// the problem

Useful tool. Wrong place for the files.

Most AI tools work by sending what you type, and the files you attach, to the provider's servers. For a firm with a duty of confidentiality, that raises questions somebody has to be able to answer: where the data goes, who can see it, how long it's kept, and whether it trains the next model.

A model that runs on your own machine takes those questions off the table. What the professional bodies say, and what the main providers' own terms say, is in Is ChatGPT safe for client data?

// what it does

Answers from your own files.

Built around one job in your firm, not a general chatbot.

  1. Questions about your documents

    "What did we agree with this client about fees?" The answer comes with the document and passage it came from, so anyone can check it in seconds.

  2. Summaries and first drafts

    Long files summarised, first drafts of routine letters, email sorted by what it needs. A person still signs off.

  3. Says when it doesn't know

    It's built to say when the answer isn't in your documents. Each answer is checked against its source before you see it, and one that fails the check is held back.

  4. Trained for your work

    Local models are open-weight: the model itself is yours to adjust. It can be fine-tuned on examples of your own work, on your own hardware, for your documents, your terms and your formats, then tested again on the same planted questions, so you can see whether it helped.

// the honest part

Smaller model, tested on your work.

  • A model that fits on your machine is smaller than the largest cloud models. It doesn't need to know everything, only your work, which is what fine-tuning is for. Whether it's good enough is not something to take on trust: before you commit, we test it on your kind of documents, with planted questions it has to get right and planted ones it has to refuse.
  • Your files stay on your machine while I build it, too. The work happens on your hardware, on site or over a remote session you start and can watch, and your documents are never copied off it.
  • It needs suitable hardware. Usually a workstation or server with a capable graphics card. I'll tell you on the call whether what you have is enough, or what to buy, before any build.
  • It's a capability I offer, not a case study yet. I build AI systems where code makes the decisions and the model only writes them up (one, shown part for part), and I test AI with traps it never sees (how). Running that on a firm's own hardware is the same work in a different place.

Tell me which job eats your team's week. I'll tell you whether a private model can take it on.

Private AI is scoped and priced on what the job needs. We talk it through on a free 15-minute call, and you get one fixed number before anything begins. See the AI offer

Start a project Book a 15-min intro call

Also on offer: custom software and automation · the AI-built app check