Promythicus / Blog
Private AI for professional firms: plain answers
We build private AI servers for professional firms. They run in the firm’s own office, so its documents never go through a cloud. These articles answer the questions we get asked.
Three ways of reading a contract, joined, beat any one of them
Every way we had a local AI model read a contract found about 82% of the clauses lawyers marked. Three of them joined by code found 87%, and 83% on long contracts where one reading found 70%. Every quote is pasted from the file, never typed.
A firm’s facts belong in a written table, not in the model’s weights
We gave a local AI model a firm’s facts in two ways. Edited into its weights, they held in only one prompt format. Written by the model as a table, after two new papers: every fact recalled in every setting, and none given out for another firm.
One AI model that searches and answers: no separate search engine
Third test on CUAD. The same local model that answers also does the search, from its own internal states and a yes/no check, with nothing trained. It quoted 83% of the clauses lawyers marked, level with reading the whole contract. The cost is speed.
We repaired an AI model by editing its numbers directly. The cleverer ideas failed.
A small open model gets stuck thinking in circles. Changing 64 neurons in one layer, with no training, took it from 7.5% to 24.5% right on maths problems it had not seen. Three more ambitious edits did nothing or made it worse.
Can a firm’s facts be written straight into an AI model’s file? Partly.
We wrote 24 facts about a fictional firm into a local model’s file with one calculation and no training. It recalled 91 to 94% of them on new phrasings, but only in the exact setting it was fitted in. We also made the model terse by adding a direction while it runs.
What is a private AI server?
A private AI server is a computer in your own office that runs an AI model locally. Your documents are read on that machine and never go through a cloud.
Where to get a private AI server in Montreal and Laval
Promythicus Inc. builds and installs private AI servers for law firms, notaries and accounting firms in Montreal, Laval and the rest of Quebec, in person, in English and French.
Law 25 and AI: can a Quebec firm put client files in a cloud AI?
Law 25 does not ban cloud AI. It asks for a privacy impact assessment before personal information leaves Quebec, and the firm stays responsible. A server in your office avoids the question.
Private AI for law firms and notaries: using AI without breaking professional secrecy
How a law firm or notary office can use AI on client files: the model runs on a server in the office, so nothing covered by professional secrecy is sent to an outside provider.
Private AI server or cloud AI (ChatGPT, Copilot): which one for a professional firm?
A plain comparison of a private AI server and cloud AI for a firm that holds confidential files: where the documents go, what each is good at, cost, and when cloud is the better choice.
Private AI servers in Ontario: Ottawa and Toronto
Promythicus Inc. installs private AI servers for law and accounting firms in Ottawa, Toronto and the rest of Ontario. We travel to your office, install the machine, and support it remotely.
Private AI servers in British Columbia: Vancouver and Victoria
Promythicus Inc., a British Columbia company, installs private AI servers for law and accounting firms in Vancouver, Victoria and the rest of BC. Documents stay on a machine in your office.
How to set up your own private AI server, step by step
A step-by-step guide to building a private AI server: the hardware, installing a model runner, choosing an open-weight model, adding a chat screen for the office, and cutting the machine off from the internet.
What hardware does a private AI server need?
How to size a private AI server: the graphics card’s memory decides which model fits. A rule of thumb for memory, with our own measurements from a 16 GB card running a 27-billion-parameter model.
Making a local AI answer from your own documents, with the page shown
Two ways to make a private AI answer from a firm’s files: search-then-read (RAG) and reading every document in turn. When each one fits, and how to make the answer show its source.
What we learned making a local model accurate on legal documents
First-hand notes from building a private document assistant: where a local model made mistakes on leases, and the four changes that took it to 16 out of 16 on our test questions at about six seconds each.
How to verify that an AI is really private
Five checks anyone can ask for to confirm an AI system sends nothing outside the office: unplug it, read the firewall rules, watch the network, list what is listening, and ask where each part runs.
How to choose an open-weight AI model for private use
How to pick a model for a private AI server: check the licence, fit it to your graphics memory, prefer 4-bit files, and test candidates on your own documents and languages before deciding.
How to secure a private AI server in an office
A checklist for securing a private AI server that holds client files: disk encryption, no outgoing internet, accounts and roles, a record of questions, backups, and controlled updates.
What does a private AI server cost in Canada?
A private AI server for a professional firm costs CA$24,500, CA$36,500 or CA$54,000, paid once, depending on the size of the office. What each price includes, what there is no fee for, and how it compares with per-user AI over three years.
Why open-source AI matters, and why private firms should support it
Open-weight AI models are the only way some organizations can use AI at all: law firms, clinics, defence suppliers, anyone whose data cannot go to a cloud. Why that matters and how to support it.
Using a local AI on scanned documents: character recognition without the cloud
How to make scanned PDFs readable by a private AI server using open-source character recognition (Tesseract, OCRmyPDF) on the same machine, what goes wrong, and how to check the result.
Can several people use a private AI server at the same time?
How one private AI server handles several users: questions are answered side by side, each taking a share of the graphics card’s memory. How to size it and what staff will notice.
The context window explained: how much can a local AI read at once?
What a context window and a token are, how many pages fit, why a longer context costs graphics memory, and what to do when a set of documents is larger than the window.
Private AI servers anywhere in Canada: how installation works
Promythicus Inc. installs private AI servers for professional firms anywhere in Canada. The demo is done remotely on your sample documents; then we build the server, travel to your office, install it and train your staff.
Search first or read the whole contract? A local AI tested on 102 contracts
We ran a private AI server on CUAD, the standard contract-review test. Reading the whole contract found 82% of the clauses lawyers marked; the best search found 71% while reading a seventh of the text. Here is what made the search better.
An outline the AI writes itself beats searching the contract
Second test on CUAD. The model labels each part of a contract in one line, then picks the parts to read from that outline. It quoted 81% of the clauses lawyers marked, against 71% for the best search and 82% for reading the whole contract.