About
For a bank, a hospital, a defense supplier or a law firm, that is not a tradeoff to weigh. It is a non-starter. So the institutions holding the most consequential information in the economy are sitting out the most useful technology in a generation. That is the thing worth fixing.
The position
Capability arrived witha condition attached.
Every meaningful jump in AI over the last three years shipped the same way: as an endpoint. To use it, you send your data to a company you do not control, running on hardware you cannot inspect, under terms you did not write. For most businesses that is a fine trade. For some, it ends the conversation in the first meeting.
It is not caution and it is not technophobia. It is the job. You cannot answer to a supervisor with a vendor’s privacy policy. You cannot put privileged material in someone else’s log. The rules these institutions operate under are not a preference they could be talked out of — they are the reason anyone trusts them with the data in the first place.
- A bank
- Customer records it answers to a regulator for, on a schedule, in writing.
- A hospital
- Patient data it is obligated to protect, under rules it does not get to reinterpret.
- A defense supplier
- Programs that are not permitted to touch a public network at all.
- A law firm
- Client material it holds under privilege and cannot hand to a third party.
The technology works for them. The delivery model does not.
Nobody is going to fix that from the outside, because the incentive runs the other way: a hosted endpoint is easier to build, easier to meter, and easier to sell. So we are building the other option — the same class of system, delivered inside the wall instead of across it.
What we believe
Three things we arenot flexible about.
- 01
An answer without a source is a rumor.
In a regulated business, an unattributed answer is not an answer — it is work. Somebody now has to go find out whether it is true. CompanyMind builds answers out of retrieved spans and cites every clause back to the artifact it came from, because an answer you cannot check is worth less than no answer at all.
- 02
Software should run where the data already lives.
Moving sensitive data is the risky part of most architectures, and most vendors resolve it by asking you to do it anyway. We think the deployment bends to the institution: your datacenter, your VPC, your air-gapped rack. Your controls already cover that boundary. We do not ask you to draw a new one around us.
- 03
Control is not the price of capability.
Choosing between capable AI and custody of your own information is an artifact of how this industry decided to ship, not a property of the technology. Run the models on your hardware and keep the index on your disks and the tradeoff stops existing. Harder to build. Not impossible.
Stage
Pre-launch.We will say so plainly.
CompanyMind is being built with a small number of regulated teams who have this problem badly enough to help us solve it properly. Everything else about our stage is on this page, in the plainest terms we can manage, because the alternative is asking you to discover it later.
Customers
None yet. There is no logo strip on this site because there is nothing honest to put in one.
Certifications
None. We hold no security certifications today and we will not imply otherwise with a badge. When we have been audited, we will say so, and you will be able to check it.
Case studies
None. Pre-launch means there is nothing deployed long enough to write one worth reading.
Revenue
None. We are not selling seats yet. We are choosing partners.
What we do have: a position we can defend line by line, and an architecture whose properties are true on day one of any deployment because they follow from how it is built rather than from how many people bought it.
The lab
A small lab,on purpose.
Sovereign software is an engineering problem long before it is a sales problem. It has to install into a building we have never entered, on hardware we do not control, with no route home and nobody from CompanyMind in the room. That work rewards a small team that can hold the whole system in its head. Below is how it divides.
Roles, not headshots. The names go up when there are real ones.
- 01name to follow
Founder / ML
position · retrieval · grounding
Owns the model layer: retrieval, grounding, and the post-training work that keeps an answer inside the source material it was built from. Decides what CompanyMind refuses to do. Also the person who replies to your first email.
- 02name to follow
Systems
deployment · inference · air gap
Owns the install. Packages CompanyMind so it lands in a datacenter we have never seen, runs inference on the customer’s own GPUs, and takes upgrades across an air gap without a support engineer standing next to the rack.
- 03name to follow
Applied research
ingestion · retrieval quality · evaluation
Works the hostile end of the problem: scans that are pictures of text, recordings nobody transcribed, spreadsheets that are secretly databases. Builds the evaluation harness that measures quality inside the customer’s boundary, because we never get to look at their data ourselves.
- 04name to follow
Security engineering
perimeter · audit trail · review
Reads our architecture the way a customer’s auditor is going to read it, then builds what that auditor will ask for: data-flow documentation, deployment topology, an audit trail that lands in the customer’s own systems. Their controls do the verifying. Our job is to leave them nothing to guess about.
Design partners
If your data cannotleave, we should talk.
We are choosing a small number of regulated teams to build this with. You get the system inside your walls and real influence over what it becomes. We get the only thing worth having before launch: the truth about whether it works.
One reply from a human. No sequence, no drip.