"Local AI" is usually sold with privacy built into the phrase, as if the address did the work. It is worth being exact, because the truth is more useful: local can be private, and so can a service — and neither is private by default. Privacy is a property of design, not of location. Something can run in your house and still quietly report back about you, and something can run far away and be built so that nothing moves without your say-so. This article explains what actually decides it, and where our own service stands, plainly.
What "local" genuinely buys
Run a model on your own machine and, for the work it does itself, no request is made: your question goes into the machine's memory, the chip does the arithmetic, and the answer comes back. That is a real property, and we will not talk it down.
But notice how little it settles on its own. The moment an assistant is useful — reading mail, watching a calendar, fetching a file — it is on the internet anyway. A box in the house can still carry software that reports back. And the box hands you its care as your jobs: the door lock, the Wi-Fi, the updates, the backups. A machine on a desk is only as private as the room, the network and the discipline around it.
Name the worry, then check the design
Privacy questions get much clearer once you name what you are worried about. Here are the usual five, each with the design question that answers it — a question you should ask of any arrangement, ours included.
- The company reading it. With a rented consumer app, every message passes through the company's computers under policies that can change. The design question is who can read what, and what is written down. In our service, your material sits in an encrypted space of your own, scoped to you alone; whatever your assistant does is logged; and answers cite your own records — so the word private is checkable rather than decorative.
- Your words becoming somebody's product. A common worry is that what you type gets used to build the next version. In our service the answer is flat: your material is never pooled, never shared, never used to train anyone else's model — and the models your assistant uses are not trained on what you send them, and what you send them is not shared with the companies that make them.
- Someone listening on the way. Anything crossing the internet can in principle be logged somewhere along the route. Everything between you and the service is encrypted in transit. And inside the service, when your assistant reaches its models, that traffic travels private networking inside our own cloud boundary — it does not cross the public internet.
- A breach somewhere else. The design questions are where the copies are and who holds the keys. Your material is encrypted at rest, with keys held in AWS KMS, and the stores that hold it have no public endpoints — there is no address on the open internet where your archive can be asked for.
- Someone getting at the machine itself. This is the honest flip side of the box on the shelf: when your work lives in your house, the lock on the door and the state of your Wi-Fi are part of the design, and they are yours. In a service, that job moves to us — and from a promise into an architecture we can describe in writing.
No arrangement is perfectly private. The design questions are how you find out what you are actually being offered — by us or by anyone else.
Where Digital Twin Pro runs, plainly
Your assistant runs in our private cloud on Amazon Web Services, in the United States. Your material lives in your own encrypted space — your files, your messages, the assistant's memory of them — with every record scoped to you, encrypted at rest and whenever it moves. When your assistant thinks, it reaches its models through Amazon Bedrock, inside the same boundary as everything else, and guardrails run on that traffic in addition to our own authorization rules. The credentials that connect your accounts are held in a secrets vault, retrieved just in time, and never shown to the models. Nothing is sent or booked without your approval; approvals expire; everything is logged. The full story is here, in plain words.
The ordinary exceptions
There are ordinary exceptions, and being straight about them is the whole point:
- It needs the internet. This is a cloud service: your questions travel to it, and without a connection you cannot reach it. We will not dress that up. Nothing is lost while you are offline — backups run continuously, encrypted — and your assistant is exactly where you left it when the connection returns.
- You are trusting us to run what we describe. That is true of every service, so we make the trust checkable: answers cite your own records; everything the assistant does is logged; what we loaded is set down in a written scope; you can ask for a complete copy of your material at any time; and corrections or deletions of what the assistant remembers are honored, confirmed in writing.
- Endings are part of privacy. When the service ends, your space and everything in it is deleted, and the deletion is confirmed to you in writing.
Frequently asked questions
Is local AI actually private?
It can be. For the work a local model does itself, no request leaves the machine, and that is a real property. But location alone settles very little: a machine in your house can still report back about you, the useful work touches the internet anyway, and the door lock, the Wi-Fi and the backups become your jobs. What makes any arrangement private is its design — what leaves without your say-so, who can read what, what is kept, and what is deleted.
Is a cloud assistant automatically less private?
No. It is exactly as private as its design. Ours runs in our private cloud on Amazon Web Services: your material in an encrypted space of your own, scoped to you alone; model traffic over private networking inside our own boundary, not the public internet; never pooled, never shared, never used to train anyone else's model; and everything the assistant does logged, with answers that cite your own records.
Do you train on my material?
No. Your material is never pooled, never shared and never used to train anyone else's model — and the models your assistant uses are not trained on what you send them. What happens instead is narrower: your archive is loaded and organized so your assistant answers from your own records, with citations, and what it learns about you serves you and nobody else. Corrections and deletions of what it remembers are honored, and confirmed in writing.
What do the models see?
Only what a given piece of work needs — the minimum context retrieved for that question — travelling over private networking inside our own cloud boundary, never the public internet. The credentials that connect your accounts are held in a secrets vault, retrieved just in time, and never shown to the models. Guardrails run on model traffic in addition to our own authorization rules, and what the models see is not used to train them and not shared with the companies that make them.
What happens to my material if I leave?
You stop in writing, at any time, and it takes effect at the end of the paid year; a year confirmed in our letter is final. Ask for a complete copy of your material and you receive it — you can do this at any time, not only on the way out. Then your space and everything in it is deleted, and the deletion is confirmed to you in writing. Nothing about you stays behind.
Your own AI. Your own material. One assistant, yours alone.
$9,995 the first year — everything included. $5,995 each year after. The first year costs more because it contains the preparation — your archive prepared and loaded by hand under a written scope, and your assistant built on it — as well as the first year of the service; every year after is the service running.
Or add the Archive Assessment first — $1,950, credited in full
You order and pay at checkout; we write back within days — a decline returns every dollar, and until our letter confirms the year you may withdraw in writing. From that letter the year is final, and it arrives on the date the letter names. Prefer to write to us first?