Offering 01 · Local & private AI
Modern AI, inside your walls.
A company-internal assistant that answers from your own documents and data — in plain language, with the source attached. It runs on infrastructure you control: from carefully approved cloud flows, up to a model that never leaves your building. We deliver the whole thing: software, integration, and the right hardware, sized to the job.
The architecture
Where your data lives should be a floor plan, not a promise.
Every AI product will promise you privacy in a policy document. We'd rather show you the construction — the drawing below is our answer to "where does our data go?"
Three configurations
As private as you need — honestly sized.
Privacy isn't one switch; it's a dial. We build all three configurations and tell you what each one can and can't do.
Approved cloud flows
The strongest models, with the smallest footprint we can design: you see and approve exactly what may leave, and nothing else does. The right call when raw capability matters more than residency.
A server in your name
Open models on a single EU server, rented in your name. Your data stays in one place you can point to, with a processor list that fits on one line.
Fully local
A model on your own hardware, in your own building — nothing leaves. We size the machine to the job and deliver it as part of the build.
Bigger cloud models are genuinely better at some things. When that's true for your job, we say so, and you decide what, if anything, they may see.
In practice
What an assistant like this actually does.
An assistant that knows your paperwork
Contracts, offers, correspondence — ask "what did we agree with X?" and get the answer with the source document instead of a folder hunt. It runs where your files live: on your own hardware, if nothing may leave the building.
Plain answers from your own database
Ask a question in plain language, get an answer from your own database: read-only, with data flows you approve, up to a model that runs entirely on your own hardware.
The colleague who has read everything
Policies, manuals, past projects — new people ask the assistant instead of interrupting the three veterans who hold it all in memory. Every answer arrives with the page it came from.
Candor
Who this is for — and who it isn't.
You're a fit if
- Your data is sensitive enough that "paste it into a public chatbot" was never an option: client files, trade secrets, personal records.
- People are quietly doing it anyway, in personal accounts — and that thought keeps you up.
- You answer to a professional duty of confidentiality: advisors, practices, anyone whose clients assume discretion.
- The same questions get answered from memory, by the same three people, all day long.
Not a fit — yet — if
- You want AI to make decisions unsupervised. We build assistants that answer and draft, with sources; judgment stays with your people.
- You mainly need AI for public-facing content. The public cloud tools are good at that, and cheaper.
- If a public AI tool with a proper business agreement genuinely fits your data, that's the cheaper answer, and our scoping will say so.
Fair questions
What buyers ask us about this.
Can a local model really be good enough?
For answering from your own documents and data — yes. Current open models run well on a single machine and are more than enough for that job. We size the model to the task; where a bigger cloud model would genuinely do better, the scoping says so. You decide what, if anything, it may see.
What hardware does this need?
Usually one machine, sized to the job. It arrives as part of the build rather than as a shopping list: software, hardware, and integration at one price.
Who sees our data while you build?
The system is built to run where your data already lives. What we may access during the build is agreed in writing before we start, and the finished system needs no outside access at all.
What happens when better models come out?
Models are replaceable parts. The integration, your data, and the workflows stay; the model is a file you swap. Open models improve fast; the machine we size today runs next year's models too.
Proof
This pattern already runs in production.
Our flagship engagement includes exactly this pattern: plain-language questions answered from the client's own database — read-only, over data flows the owner approved. It runs today, 24/7, inside a system the client owns outright. The same architecture, turned further toward privacy, is what Drawing № 01 shows: everything inside the walls.
That's one engagement, told honestly. We'd rather prove a pattern once, deeply, than gesture at a dozen; the dial goes as far as your data requires.
How to start
Where may your data not go?
Tell us what you'd ask an assistant if you could trust one — and where the data lives that would answer it. The scoping conversation is free and binds you to nothing: you'll get a read on which configuration fits, or whether you need this at all.