What this is
Inference you can point at
Serving, identity, keys, backups, a surface your team talks to, and a one-page Admin Card. This is not the company brain — no map, no morning brief, no ingest pack. Those are a separate install.
- Container host in the environment you choose.
- Model-serving stack with health checks. Silicon is a slot — swap later without redoing the box.
- Sign-in against the Google or Microsoft you already have. No new identity provider.
- Credentials minted in your secret store. We never hold root.
- Encrypted backup you control, and a restore drill before we leave.
You pick · we recommend · always your environment
Three pictures of ownership
In your tenancy
Workhorse model as reserved or scale-to-zero GPUs in your GCP, Azure, or AWS account. Default when there's no office box and you already have a cloud bill.
On the LAN + in your VPC
Everyday work on hardware that stays in your office, with a batch model in your cloud. Same ownership picture. Two places the work can run.
Every cable is yours
Local cluster, local adjudicator, no required egress. For teams who want to show every cable, or have no usable cloud tenancy.
What this is not
Not a wrapper around a lab
A vendor API with the lights off is not private AI.
We will not sell a “private GPT” that still sends your work to a lab. If you later insist on a frontier burst, that’s a written exception with an egress receipt — not the product.
Also not included: a knowledge map, connectors, a morning brief, or “your Drive answers now.” Those ship with the company brain install.
What you own when we leave
The stack stays. The Admin Card names how to revoke us.
If you walk, serving stays yours. Same cancel-and-keep spine as the brain install, scoped to inference. built on Mnemma
Next step
See which topology fits.
To get started, book a fit call. We'll map cloud, hybrid, or on-prem against the environment you actually have.