Trust center
Security, privacy and AI governance.
Where your data lives, who can read it, what the AI may and may not do, and what we do not have yet. One link for your security review.
- 01 | On every requestA permission check
- 02 | In every queryAn organization filter
- 03 | For background workA per-organization context
- 04 | Everywhere elseDeny-by-default policies
- 05 | In the database itselfRow-level security, forced onA query that forgets the organization filter gets no rows
Four facts first. One of them is a gap.
- Stored in CanadaMontréal region, with backups in the same region.In place
- Isolated in the databaseRow-level security on every organization’s tables.In place
- A person decidesAI proposes. Humans approve. No AI can move a record’s workflow stage.In place
- SOC 2 Type IIn preparation. No certification is claimed today.In preparation
Five layers between one customer and the next.
A query that forgets the organization filter gets no rows: row-level security is forced on every organization’s tables, so a missing filter fails closed.
- A permission checkOn every requestIn place
- An organization filterIn every queryIn place
- A per-organization contextFor background workIn place
- Deny-by-default policiesEverywhere elseIn place
- Row-level security, forced onIn the database itselfIn place
Stored in Montréal. Every stop is named.
Point-in-time recovery is not enabled today, so up to a day of changes could be lost in a full restore.
- Stored
- Montréal, Canada
- Application
- United States, US East
- AI requests
- OpenRouter
- Files and backups
- The storage region
- At rest
- AES-256
- In transit
- HTTPS, TLS 1.0 and 1.1 refused
Four documents. One review.
What a security review asks. Answered.
Five layers: a permission check on every request, an organization filter in every query, a per-organization context for background work, deny-by-default policies, and row-level security in the database itself, forced on under a role that cannot bypass it.
The database returns no rows. Row-level security is forced on every organization’s tables, so a missing filter fails closed.
A few named platform jobs do, such as the opt-in benchmark, which builds de-identified figures from at least ten operators. They run under a separate role and never show one operator’s records to another.
In Canada, in the Montréal region of our database provider. Files and backups stay in the same region.
The application runs in the United States, in the US East region. AI requests go to OpenRouter. Every sub-processor is named in the register, with its role and region.
AES-256 at rest on the database, files and backups, as attested by our providers. HTTPS everywhere, with strict transport security; TLS 1.0 and 1.1 are refused.
Seven rolling daily backups, in the same region. Point-in-time recovery is not enabled today, so up to a day of changes could be lost in a full restore.
Yes, at any time: every record and the audit log, as JSON or CSV, delivered as a signed download link that lasts seven days. Occurrences also export as E5X for ECCAIRS.
Inside eAviora, eAvy proposes and a person accepts, changes or dismisses before anything is written. One exception, started by you: a bulk import can apply high-confidence classifications, marked unvalidated, and sends the rest and a sample to a person. An outside assistant you connect over MCP acts with that person’s own role, and its writes are audited. No AI can move a record’s workflow stage.
Only the record fields an agent is set up to read. A confidential or anonymous reporter is never part of it.
eAviora trains no model on your data. AI requests are processed through OpenRouter; ask us how the model providers behind it handle requests, and we will show you the current setting.
In credits, from one shared monthly pool per organization, with hard caps you set.
Yes. The name of a confidential reporter shows only to people cleared for confidential reports, on the record and in its audit, timeline and export views. On an anonymous report, the reporter’s name is hidden from everyone, our own staff included.
Each record carries one of three levels: standard, confidential or privileged. Without clearance it does not appear in lists or notifications, and the assistant does not show its content.
Tail numbers and dates are removed automatically, the draft is written without names, and a person approves the bulletin before it is shared.
Yes: SAML 2.0 with Okta, Microsoft Entra or Google, accounts created at first sign-in, and single sign-on enforced for your organization if you choose.
Yes, through SCIM 2.0, so a leaver loses access without a manual step.
Authenticator codes for every user, with a fresh check before sensitive actions. Passkeys are available as an experimental option.
Yes: active sessions, the last 20 sign-ins, and sign out everywhere in one step.
Each key has a read, write or admin scope, is stored hashed, is shown once when created, and records when it was last used.
Every change made in the app or through the API writes its audit entry in the same transaction as the change. Sign-in and provisioning events are logged too, outside that transaction.
Updates are refused by the database itself, and so are deletes, outside one gated maintenance path. Entries are chained, so a gap shows.
Yes: search by record, person, date and action in the app, and export it as JSON or CSV.
Not yet SOC 2 Type I is in preparation. No certification is claimed today. Type II follows the required observation period.
Not yet Not yet. No third-party penetration test has been performed, so there is no report to share.
Not yet No. Our hosting providers hold their own certifications; eAviora does not.
Not yet Not today. Keys are managed by our database provider.
Walk your security team through it.
Bring your questionnaire. We answer it live, on the running platform.