Does any of our data leave the network — including for inference?

No. In an on-premise or private-cloud deployment, prompts, documents and embeddings never cross the boundary, and inference runs on your own GPUs. The one exception is deliberate: route an agent to a hosted model provider, and its calls go where you point them. Leave hosted providers out and there is no exception at all.

We are not one site. Can one deployment cover several hospitals, departments or regions?

Both shapes are normal. One install can serve every department: groups, agents and filestores are tenant-scoped, so no department reads a corpus it was not granted. Where a residency rule pins data to one place, run a separate install there, with its own keys and audit log. The technical assessment settles which shape fits.

Can it run with no internet connection at all?

Yes. Air-gapped is a supported deployment target, not a special case. Models, platform updates and security patches arrive as signed bundles applied on your schedule through a controlled transfer, and nothing travels the other way — no telemetry, no usage reporting, no error collection. Day-to-day operation needs no route out.

Do we give up model quality by keeping inference inside?

You choose per agent. Open-weight models on your own GPUs handle planning, retrieval, extraction and drafting: the bulk of the work. Where one agent genuinely needs a hosted frontier model and its inputs are not sensitive, route it out; the rest stay local. Most deployments run both; an air-gapped site runs the local half.

What hardware do we need?

That comes out of the technical assessment rather than a datasheet: it depends on your document volume, how many agents run concurrently, and which models you serve locally. The platform scales horizontally as volume grows. We would rather size it with you than publish a minimum that reads as a quote.

Who operates it after go-live?

Your team. We install, integrate, run the security review, commission it alongside your engineers and hand over the runbooks — the deployment keeps working whether or not we are engaged. Ongoing support and an evolution track are available as a contract you choose, not a dependency built into the architecture.

Is our data used to improve your models?

No, and in an on-premise deployment there is no mechanism by which it could be — there is no pipeline out of your network for it to travel down. Prompts, documents and corrections stay in your storage, under your keys, governed by your retention policy.

Sovereign deployment

Nothing leaves the building.

The whole platform installs inside your network — agents, retrieval, extraction, audit, and the inference itself. Bare metal, private cloud, or fully air-gapped.

Talk to our deployment teamScroll the drawing.

100% on-premise
Air-gap capable
Inference on your GPUs
SOC 2 · HIPAA · ISO 27001

Sheet 01 — on-premise deployment · plan + rack elevation01 / 07

One line, and it does not move.

Everything on the left belongs to someone else's network — the model vendors, the public internet, and us. Everything on the right is yours.

The perimeter

The same install. You choose where the line sits.

Nothing about the platform changes between these — same units, same governance, same console. What changes is where the boundary is drawn and who is standing inside it.

Bare metal

Your racks, your data centre.

Installed on hardware you already own and already audit. GPU sizing comes out of the technical assessment, against your real document volume.

Your hardware
Your data centre
Your network policy

Private cloud

Your account, your keys.

The same deployment inside your own VPC. No shared control plane, no multi-customer inference, nothing routed through infrastructure we operate.

Your cloud account
Your KMS keys
No shared control plane

Air-gapped

No route to the internet at all.

For sites where the network genuinely does not leave the room. Models and updates arrive as signed offline transfers; day-to-day operation needs no connection.

No egress
Signed offline updates
Local models only

Managed cloud

When sovereignty isn't the constraint.

We run it for you on dedicated, tenant-isolated infrastructure. Honest framing: this is the option to pick when residency is not what you are buying.

We operate it
Dedicated tenancy
Same governance model

The same install, whatever you point it at

The install

Five units, and none of them phone home.

This is the whole platform. Every capability the rest of the site describes — agents, skills, knowledge graph, chatbots, analytics — is one of these units doing its job inside your network.

Rack — front elevation5U

Agent runtime

The orchestrator every agent runs on: planning, tool calls, skills loaded on demand, automations, and the human-in-the-loop stops you define.

Inference — GPU

Local models served on your own accelerators through Ollama or vLLM. Route to hosted providers if you want them; leave them out entirely if you don't.

Vector index

Embeddings and retrieval over your indexed documents. Text is embedded on-site — nothing is shipped out to be turned into a vector.

OCR & extraction

Scanned documents turned into structured, cited fields on your own hardware. Claim files, referral packets, tender bundles and case papers are read where they are already stored, rather than uploaded somewhere to be read.

Audit log

Every prompt, tool call, and document access — attributable to a user, exportable, and yours. The thing a regulator asks for that most deployments cannot produce.

All units inside the boundary

What the runtime actually runs

What the boundary buys you

The documents that could never leave are the ones this reads.

Every organisation has a corpus models were never allowed near, for one reason: it could not leave. Patient records, case files, ledgers, tender bundles, the scanned decades in the basement. With the model inside the boundary, that corpus is simply readable. The question is no longer which documents may leave, but which department goes first.

The restricted material goes first, not last. The corpus that waited for a policy is the one this was installed to read.

The second use case costs a configuration. Same install, same boundary, same audit log: no new security review to add a department.

Nothing switches off when the contract does. The install keeps running on your hardware, under your keys, with or without us.

How we get there

Five revisions, and then it's yours.

The engagement is the drawing being made. Each pass adds detail against your constraints, and the last one signs it over.

REV A

Discover

A free technical assessment: your workflows, your document types, the compliance ground you stand on, and what the platform will have to integrate with.

REV B

Design

The architecture drawn against your constraints — where it sits, what it touches, which models run locally, and who is allowed to reach what.

REV C

Build

A proof of concept on your actual documents, running on your own infrastructure. Not a sandbox with sample data.

REV D

Deploy

Production install, integration testing, security review and compliance validation — then commissioning with your team in the room.

REV E

Evolve

New use cases, model upgrades, tuning as volume grows. The drawing gets revised. It stays yours.

The handover

We deploy it, then we give it to you.

The deployment is not a tenancy we rent you access to. It is an installation on your hardware that keeps working whether or not we are in the room.

Your keys, your policy

Encryption keys, IAM policies and network rules are yours from commissioning onward. There is no vendor key that opens your deployment.

Your data is never training data

Nothing in your deployment is used to train anything: not prompts, not documents, not corrections. There is no pipeline out of your network for it to travel down.

Your team operates it

We commission it and hand over the runbooks. Support is a contract you choose, not a dependency you inherit.

SOC 2 Type IIAICPA

HIPAACompliant

GDPRReady

ISO 27001Certified

How access and policy are governed

At a glance

The deployment, in specifications.

Deployment targets: Bare metal · private cloud · air-gapped · managed
Topology: One for the organisation, or one per site or jurisdiction
Inference: On-site GPU — local models via Ollama or vLLM
Model choice: Local, hosted, or both — routed per agent
Encryption: AES-256 at rest · TLS 1.3 in transit
Access control: Role-based IAM; deny overrides allow
Isolation: Per-tenant across groups, agents and filestores
Audit: Prompts, tool calls and document access
Updates: Signed transfer — no internet connection required
Languages: 120+
Certifications: SOC 2 Type II · HIPAA · GDPR-ready · ISO 27001

Questions

What buyers ask before the assessment.

Issued

Let's draw yours.

It starts with a free technical assessment — your workflows, your document types, your compliance ground — and ends with a proof of concept running on your own documents, on your own infrastructure.

Book a demo

Exemplary AI — deployment

Project: On-premise deployment
Scope: Platform, models, audit
Issued to: Your organisation
Keys held by: You
Status: Handed over

The documents that could never leave are the ones this reads.