Security and data handling

1. Where it runs

Each client gets their own deployment and their own PostgreSQL database, hosted in the Netherlands (Railway, europe-west4). The language models run on Microsoft Azure OpenAI in an EU resource, so document text and answers are stored at rest inside the EU. Those model deployments are Global Standard, which means an individual request may be processed in an Azure region outside the EU or EEA while nothing is persisted there; a client whose policy requires processing to stay inside the EU can be moved to an EU Data Zone deployment on request.

2. Separation between clients

One deployment and one database per client, with its own credentials. There is no shared environment and no query that spans clients, because there is no connection between them to query across. The website and the assistant are separate; a visitor to this site reaches no client's data.

3. Who can reach what

Signing in uses the client's own identity provider over OIDC, so their password policy, multi-factor authentication and conditional access apply unchanged, and we never hold a password. Accounts fall into three levels: a user asks questions and translates documents; an administrator also manages the document set and sees usage totals; a named superuser also sees stored conversations, the feedback queue and the quality measurements. Every level is enforced on the server for every request, not by hiding buttons. Restricting individual documents to a group within the client's organisation is available on request and is set up per deployment; by default everyone who can sign in can search the whole document set, so a deployment that holds material not meant for all staff should ask for it.

4. What is kept, and for how long

Documents are kept until removed by an administrator, or deleted with the deployment when an agreement ends. Questions and answers are retained, and so are problem reports staff file about an answer, because together they are the only record of where the assistant answers badly and what we need to fix; automatic deletion after a fixed period can be switched on per client for both, and is a term of the agreement rather than a setting we change unilaterally. Sessions expire after seven days and expired ones are deleted daily. Encrypted database backups are kept for up to 89 days, so material deleted from the live system can persist in a backup for that period. Files uploaded to the translator are converted and removed immediately; nothing is retained from them.

5. How the assistant handles a document

A document is split into passages, and only the passages relevant to a question are given to the model, each labelled with the file and page it came from so an answer can be checked against the source. Instructions written inside a document are treated as text rather than obeyed, and that is tested rather than assumed. Model providers are not permitted to train on this content. Every answer carries its citations so a reader can verify it, because a confident wrong answer is the failure mode that matters.

6. Access, correction and deletion

A signed-in user can delete their own questions and answers from within the assistant, in one action. For anything wider - a copy of what is held, a correction, or deletion for a person who has left - write to the address below and we will act within 30 days; where the client is the controller we act on their instruction. Removing a client's documents and database entirely takes one working day, and we confirm in writing when it is done.

7. If something goes wrong

Report it to the address below; we aim to acknowledge within one working day. We investigate, contain, and tell the affected client what happened, what data was involved and what we changed - within 72 hours of establishing that personal data was affected, so the client can meet their own notification duty. Where a credential may have been exposed we rotate it first and investigate second. We will say what we do not yet know rather than wait until the picture is complete.

8. What we do not claim

Udamon is not ISO 27001 certified and has not been penetration tested by a third party. Continuous security monitoring is not in place yet, though uptime is monitored and administrative actions are recorded. Registration as a legal entity is deliberately timed to the first client, so a data processing agreement is signed as part of that first engagement rather than before there is one; a draft is available on request now. We would rather a reviewer read this section than discover it.

9. Who else touches the data

The complete list. Anything added here is a change we tell clients about before it takes effect.

No analytics, advertising or tracking service is used on this website or in the assistant. No third party receives document content other than the model provider above.

10. Security contact

For a vulnerability report, a security questionnaire, a data processing agreement or a deletion request. Please include enough detail to reproduce a problem; we will not pursue anyone who reports one in good faith.