Everyone else rents a brain.
Mine owns one.
A private AI mind that lives on my own server, trained on the operating record of my own AI company. Ask it anything about the business and the answer comes back with the source underneath it.
When does the monthly retraining run, and what does it actually do?
Retraining runs on the first day of each month at 03:17 as a scheduled job.[1] It builds a fresh training set from the company's current memory, trains a candidate, and evaluates that candidate against the model currently in service. The new version is deployed only if it wins the evaluation.[2]
A rented brain is brilliant and blank
Commercial assistants know everything about the world and nothing about your company. That gap is the whole problem — and it is not solved by a better prompt.
Renting a general model
- Knows the internet. Knows nothing about your last twelve months.
- Your documents travel to somebody else's infrastructure to be read.
- Confident answers with no way to check where they came from.
- The vendor changes the model; your answers change with it.
- Access lasts exactly as long as the agreement does.
Owning a private mind
- Knows your operation — every decision, task and report that was written down.
- Runs on your machine. Nothing is sent out to be read.
- Every sentence carries the file it came from. Verifiable.
- The weights are yours. Nobody can change or withdraw them.
- It answers tomorrow exactly as it answers today. Sovereign.
How one question is answered
Think of a perfect archivist and a fluent librarian. The archivist knows where everything is and never speaks. The librarian reads beautifully but knows nothing until handed the pages.
Someone asks
In plain language — from chat, from the command center, or out loud. No syntax, no query language.
The archive is searched
Thousands of fragments of company record are ranked, de-duplicated and narrowed to the few that actually matter.
The mind reads
It receives only those excerpts, with one instruction: answer from these, or say you don't know.
An answer with receipts
Back comes a short answer and the files it stands on — so anyone can check it in seconds.
The design decision that makes it work
Most attempts to build a company model try to teach the facts into the model. That is the hard way to be wrong by Tuesday.
Kept in a living index
Every document, decision and log entry sits in a searchable index on the server. When something changes, the index rebuilds itself in about two minutes and the next answer is already current.
Nothing has to be retrained for a fact to be known.
Trained on how to answer
The model was fine-tuned only on manner: answer briefly, in the company's own vocabulary, cite a source in every sentence, and refuse rather than guess.
That is what one monthly training run maintains — tone, not truth.
Facts outside, style inside. That single separation is why the mind is never out of date, why a month of new history needs only a single training run, and why a wrong answer is visible rather than hidden.
Three places to think, one answer
The mind needs computing power to think. It is available in three places, and the system falls down the ladder on its own — so an answer always arrives.
Built in four phases — including the failures
I am publishing those too, because a build story without them is marketing, not engineering.
Prove it is worth doing
Fifteen questions, including two traps about things that were never written down, run identically on the plain server and on a graphics machine. Both passed the traps without inventing anything. Four questions came back as a false "I don't know" — which exposed the retrieval, not the model, as the weak part.
Fix what the pilot exposed
Better search repaired all four failures. It also uncovered a silent defect: the company's largest journal file had been quietly skipped by the indexer since the beginning. Repairing it added over 1,300 fragments of history the system had never been able to see.
Teach it the house style
A larger teacher model produced a training set from the real record. The first attempt over-learned and started refusing questions it knew — so it was discarded. The second, dialled to 40% strength, scored a perfect run on the evaluation set with a citation in every sentence.
Make it wake on demand
The mind moved onto infrastructure that sleeps when unused and wakes the instant a question arrives, with the company's own server underneath as a permanent fallback. Verified end to end: a question asked from a phone woke a sleeping machine and returned a cited answer, sources attached.
What this means for a company like yours
The interesting part is not that a private mind is possible. It is that it is now within reach of a company that is not a laboratory.
Institutional memory that answers
The reason a decision was made two years ago, the supplier issue nobody remembers, the audit evidence buried in a folder — asked in a sentence, answered with the document attached.
Confidential by construction
The knowledge never leaves your infrastructure to be read. For regulated environments this is not a policy promise; it is an architectural fact you can point an auditor at.
Knowledge that survives people
When someone retires or moves on, what they wrote down stays answerable. The record stops being an archive nobody opens and becomes something the company can actually ask.
No vendor holding the keys
The weights sit in your own storage. If every provider changed their terms tomorrow, the system would answer the next question exactly as it did today.
What it deliberately will not do
It answers from what has been written down — nothing more. Ask about something that was never recorded and it will tell you plainly that it doesn't have it, rather than produce a plausible sentence. It is short, factual and sourced, not a conversationalist.
That restraint is the point. A mind that would rather say "I don't know" is the only kind worth putting near a real decision.
Questions I get asked
How is this different from ChatGPT or Claude?
A commercial assistant is a general brain you rent — brilliant at language, blank about your company. Velin Sovereign is an open model running on your own server, connected to your company's own record: its journal, tasks, documents and decisions. It answers about your business specifically, and names the file each statement came from.
Is our confidential data used to train it? Where does it go?
The facts are never sent anywhere. They stay in a searchable index on your own machine, and only the few most relevant excerpts are handed to the model at the moment a question is asked. The model itself is trained on style alone. Training material never leaves the server, and the resulting weights are stored privately.
What does the model actually run on?
Your own infrastructure. The everyday path is a machine that wakes itself the moment a question arrives and sleeps again shortly after the last one, so nothing is held running for the sake of being ready. Underneath sits a permanent fallback on the company's own server — slower, but always available, so an answer arrives even if everything above it is not.
How does it stay current when the business changes daily?
Facts are kept outside the model on purpose. When a document, task or entry changes, the index rebuilds within about two minutes and the next answer already reflects it. The model itself is retrained only monthly, and only for tone.
What stops it from making things up?
Three things: it is only ever given retrieved excerpts, never asked to answer from memory; it is trained to attach a source to each statement, so an unsupported claim is visibly missing its citation; and it is tested with deliberate traps about things that were never recorded. In the acceptance run it answered every trap with "I don't have that in memory."
Can this be built for our company?
Yes — as an extension of a Velin build. The prerequisite is a written record the model can read: documents, tickets, reports, decisions. If that record exists, the private mind sits on top of it. If it doesn't exist yet, Velin creates it first and the mind follows.
Ask my company's mind a question. Live.
Send me something you'd want your own system to answer. I'll put it to the private model on a screen share, and you'll see the answer and its sources arrive together — from a machine that was asleep a minute earlier.
Book a 15-min live session