Platform · Security and AI model governance
Your data in its own instance. No third-party AI service sees it.
Each customer runs in a single-tenant instance that we host. The platform is designed to run with a local AI model, in your environment or ours. Every deployment comes with a model governance statement your model risk function and your examiner can read.
The problem it removes
AI in a compliance program is a model risk question
- Most AI tools send your customer records to a third-party model you can't inspect, and your vendor review has to account for it.
- Your model risk function is asked to approve a model that no one can document, because it changes without notice.
- Shared, multi-customer software means your data sits beside someone else's, and access rules are the only thing between them.
What you see
A governance file for your deployment
Each deployment has a model card, a controls matrix and an EU AI Act crosswalk, gathered in one model governance statement. It names the model, what it's allowed to do, and the controls around it.
| Item | Entry |
|---|---|
| Tenancy | Single-tenant instance, container, hosted by Rupture Labs |
| Language model | Open-weight model, run locally on our hardware; no fine-tuning |
| Model role | Drafts text only. Makes no determination. |
| Data to third-party AI services | None |
| Access roles | Defined roles; monitoring read-only role cannot change records |
| Attached | Model card · Controls matrix · EU AI Act crosswalk |
Illustrative example with invented entries. The statement for your deployment is written for it and signed.
How it works
Rules decide, the model writes, a person signs
- Isolate. Your data lives in its own instance, packaged as a container and hosted by us. A request for another customer's record returns "not found", never "forbidden", so the system doesn't even confirm it exists.
- Control access. Users hold defined roles. In transaction monitoring, a read-only role can never change anything.
- Decide by rule. Every determination is made by written rules. The language model only writes: summaries, narratives, first drafts. It's a standard open-weight model run locally, not a proprietary fine-tuned one, so it can be documented and reviewed like any other model.
- Review and sign. A practitioner reviews the output and signs it. The governance statement records how each of these steps is controlled.
Guardrails
What it will never do
- We don't train a proprietary model on your data. The model is standard and documented, which is the point.
- It doesn't mix your records with another customer's. Each customer runs on its own instance.
Where it shows up
The work this part does for you
Plain English
What this is, and how anyone does it
Reference articles from our library, cited to the published rules and standards. No sales copy.
Connected parts
What it works with
Talk to a practitioner
Book a 15-minute chat with our founder.
A real conversation with a senior compliance leader, to see if there's a fit. Not a sales call, not a demo, no pressure.