Sovereign by architecture
Weights, prompts, context and outputs stay on hardware the organisation owns, in a jurisdiction it chooses. Nothing leaves the boundary because the system has no route for it to leave by.
Build · Private AI Systems
The model stays where the data belongs.
Inference hardware we set up and support, and agentic harnesses that automate the work, inside a boundary the organisation owns.
Overview
We build and operate AI systems for organisations that cannot send sensitive work to a model behind a public endpoint. That takes two things. Inference hardware sized, installed and supported where the data already lives. And an agentic harness: the tools, permissions and review steps that let a model complete real work in the business, under controls the business sets. Model, hardware, data path and access boundary are specified together, so nothing leaves the agreed boundary.
What it includes
The work, named plainly.
Hardware set up and supported
We specify, procure, install and run dedicated inference hardware, from a single workstation to a rack. Capacity, firmware, model updates and failures are ours to handle, with a named engineer on call.
Models chosen on evidence
Open-weight and purpose-trained models evaluated against the real task. We measure accuracy, refusal and failure modes before choosing scale.
Agentic harnesses that do the work
A model on its own answers questions. A harness gives it the tools, data access and permissions to finish a task: triage a mailbox, reconcile a ledger, draft a record for sign-off. We build the harness around your systems, with every action scoped and logged.
A person remains accountable
Actions that affect records, decisions or people pass through explicit review. The audit trail shows what the model did and who accepted it.
Proof small enough to inspect
We designed and trained our own 12.2 million parameter language model. It runs entirely on the reader’s device.
Run the modelHow we work
Engineered once. Maintained indefinitely.
- Classify the data and define where it is allowed to move
- Measure the task before selecting a model
- Size, install and harden the inference hardware
- Build the harness: tools, permissions, review points and logging
- Evaluate accuracy, refusal and failure against written cases
- Deploy with human review and a controlled rollback
- Support the hardware and re-approve material model changes
Speak with us
Tell us where the data must stay.
We reply within one working day, from an engineer who understands the deployment.
Send securely