Security and governance
Nothing leaves the network. Here is how that is enforced.
Inference happens on your machine. There is no call out to a model provider, so the processor, the transfer, and the retention question all disappear at the architecture level rather than in a policy document.
For your security team
You are probably not the person who has to approve this.
Whoever has to sign it off will want the architecture, the data flow, and the governance answers in one place they can read without you in the room. This page is that document. Print it to PDF and send it on.
Governance
The questions an auditor opens with.
We are not your compliance officer. We build the architecture that makes these answers short.
Who processes personal data, and where?
You do, on your own infrastructure. No third-party processor and no international transfer to assess.
Can you document and control the system you deploy?
Model version, weights, prompts, and retrieval sources are all pinned and versioned in your repository.
Does patient or client data stay in the estate?
It never leaves. Air-gapped deployment removes the network path entirely, not just the policy.
What happens if the supplier disappears?
Open weights, open components, documented configuration, and your team holds the runbooks.
Reference architecture
What we actually install.
No black box. Every layer is an open component your team can read, audit, and operate without us.
Access
Web UI and an OpenAI-compatible endpoint behind your own SSO
- OIDC / SAML
- Reverse proxy
- Per-team API keys
Orchestration
Retrieval, prompt assembly, tool calls, and per-request policy
- RAG pipeline
- Role-based scoping
- Full query audit log
Knowledge
Your documents, chunked and embedded, indexed on your own storage
- Vector store
- Connectors
- Incremental sync
Inference
Open-weight models served locally, pinned to a version you approve
- Model server
- GPU or unified memory
- Version pinning
Operations
Metrics, logs, alerting, and backups wired into what you already run
- Prometheus / Grafana
- Log shipping
- Restore drills
Deployed with Docker and systemd on Linux, or natively on Apple silicon. Configuration lives in your repo, not in our heads.
The decision on one page
Same capability. Different threat model.
This is the table your security officer will build anyway. Here it is up front.
| Criterion | Cloud AI | SRI |
|---|---|---|
| Where inference runs | Vendor infrastructure, region of their choosing | Hardware you own or rent in your own name |
| Who holds your documents | A third party, under their retention policy | You. The index never leaves your disk |
| Processor agreement needed | Yes, plus transfer assessment | No third-party processor in the request path |
| Works with no internet | No | Yes, on-premise and air-gapped |
| Audit log location | Vendor console, exportable at best | Your own log stack, queryable like any other service |
| Model choice | Whatever the vendor ships and deprecates | Open weights you pin, upgrade, or roll back |
| Cost shape | Per token, unbounded, scales with success | Fixed hardware or fixed monthly, flat under load |
| If the relationship ends | Access stops, data export on their terms | The system keeps running. You already own it |
Next step
Tell us what your team is not allowed to do yet.
A 45-minute call. We look at your workflows, your data rules, and whether local AI is worth it for you. If it is not, we will tell you that instead of selling you a project.
Not ready for a call? Send the security brief to whoever has to approve it, or just reply to an email. All three work.