Deployment & trust
Your infrastructure. Your models. Your rules.
Every AI vendor gets asked where the documents go. With Litmus, you decide the answer, and your security team can check it.
No internet connection. The language model and every other component run on your network.
Three ways to deploy
The same software, wherever it needs to run.
What runs where
Component by component.
What changes between the three models is who runs each part, and which services it calls.
| Component | On your infrastructure | Managed by Litmus | Air-gapped |
|---|---|---|---|
| Application and database | Your server. | An installation hosted and operated by Litmus. Hosting details are agreed during scoping. | Your server, on a network with no internet access. |
| Search and re-ranking | Open-weight models run inside the application, on the same server. In every model, search sends no document text to a third party. | ||
| Language model | Any OpenAI-compatible endpoint you choose, cloud or self-hosted, with your key. | Your endpoint and key if your administrator sets them. Otherwise, Litmus’s provider account. | A model you host on your network. |
| Document reading (OCR) | A supported cloud OCR service, with your key. | The same choice, with your key or Litmus’s. | Runs on-premises. |
| Speech-to-text for minutes | A supported cloud speech service with your key, or a transcription endpoint you run. | Your key if you set one. Otherwise, Litmus’s provider account. | A transcription engine on your network. |
| Updates | Signed releases, applied by your administrator. | Applied by Litmus across the service. | Signed offline bundles, uploaded by your administrator. |
Litmus does not train or fine-tune any model on your documents. The model, OCR and speech providers are your choice, and with your own keys they process content under your terms with them.
Designed for scrutiny
Built expecting to be audited.
The model is given no tools.
Litmus sends the model text and receives text back. The model cannot send, change or delete anything, and what it writes stays a draft until an auditor accepts it.
Access is checked before the prompt is built.
Passages are filtered by organisation, sensitivity level and scope before the model sees them. A manipulated document can change what the model says. It cannot change what the model was shown.
Roles, sensitivity levels and scoping.
Five built-in roles, plus custom roles set per module. Four sensitivity levels you can rename, with a ceiling on each role. Scoped users see only the engagements they are staffed on.
An administrator who can’t read the audit.
The IT Administrator role manages users, roles, assignments and the licence, and cannot open any finding, report or document.
Your identity provider, and its MFA.
Microsoft Entra ID and Google sign-in can be enabled for your organisation, with MFA from your identity provider. Accounts are set up through one-time links, so administrators don’t need to know passwords.
Your model, your keys.
Any OpenAI-compatible endpoint, cloud or self-hosted. Stored provider keys are encrypted under a per-organisation key, and backups are encrypted.
Signed updates.
Each release is verified against Litmus’s signing key before it is loaded, and a backup is taken first.
Error reports carry no document text.
Error reporting is off unless you enable it. Each report is rebuilt from an allow-list, with no request bodies, and email addresses, file paths and document names are redacted.
Every request can be traced.
Each response carries a request ID, and every log line records it with the organisation, so a user’s action can be traced to the system’s response.
For your CISO and IT architects
The two-page IT & security brief.
The detail your security team will ask for, component by component, written to be forwarded. Tell us where to send it and we’ll email it to you.
- Three ways to deploy
- What runs where
- What leaves your environment
- Identity and access
- How the AI is governed
- Operations, export and deletion
- Questions to ask any AI audit vendor
Straight answers
Ask us the hard questions.
Where does the data live?
Wherever you decide. On your infrastructure, the application and database run on servers you own and administer, and no application port is opened to the internet. Managed, Litmus hosts and operates the installation, with hosting details agreed during scoping. Air-gapped, everything runs on a network with no internet connection, including the language model.
For audit and risk teams in the UAE, this is usually the first question asked of any AI vendor, and it should be answered before anything else is discussed. We go further in air-gapped AI and data sovereignty for UAE audit teams.
Who can see it?
Only the people your roles allow. Access is set by role, sensitivity level and engagement or department, and it is checked before any content reaches the model, so an answer draws only on what that user may see. The IT Administrator role runs the platform and cannot open audit content. Litmus does not train models on your documents.
What happens when we leave?
On your infrastructure or air-gapped, the data is on your servers, so there is nothing to retrieve from us and removing it is your process. In the managed model, Litmus exports your organisation’s data as one bundle on request and deletes it at the end of the service, with a waiting period before deletion is final. Backups expire on their normal schedule. Reports export to .docx and committee decks to .pptx.
The litmus test for trust.
Bring your security team to the demo.
Their questions are welcome from the first meeting. Ask for the IT & security brief to share with them ahead of it.