Security and data handling at ceadr

ceadr is a multi-tenant SaaS. The web application runs on Vercel, the API engine and background workers run on Railway, and data lives in managed Postgres and object storage on Supabase. Authentication is handled by Clerk.

Each workspace's data is isolated by Postgres row-level security and re-pinned to the workspace at the application layer on every read and write. One customer's content is never visible to another.

All traffic to the application and API is served over TLS. The managed Postgres database and object storage are encrypted at rest, and secrets and connection credentials are additionally encrypted at the application layer.

To make your knowledge queryable, connected and uploaded content is split into chunks, embedded as vectors, and stored so it can be retrieved and cited. A durable knowledge graph of entities, relationships, and cited excerpts is built from it. This is a persistent knowledge platform, not an ephemeral scan-and-discard tool: the persistence is what lets an agent answer a question with a citation next week.

We do not train on your data. Content sent to Anthropic's Claude API is not used by Anthropic to train models. We do not train our own models on your content, and we do not sell your data.

Nango holds the keys for connections made in the app. Connections you make in the ceadr app are brokered by Nango, so ceadr stores a connection reference rather than your provider access token. An older direct connection route still exists; tokens saved through it, including any refresh token, are kept encrypted in our own cache for 24 hours after they were last saved.

Read by default. ceadr only reads from your connected sources unless you explicitly approve a writeback, and write access is disabled per connector until you enable it.

Your data persists while your workspace is active. You can export it, permanently erase it, or delete the whole workspace at any time. We honour data subject rights under GDPR Articles 15 to 22 (access, rectification, erasure, portability, restriction, objection), the UK DPA 2018, and CCPA.

To exercise a right, email julian@ceadr.ai; we respond within 30 days. Workspace admins can also open export and erasure requests directly from workspace settings.

Owner, admin, member, and viewer roles are enforced at the knowledge layer, not just at the document level. Every privileged action is written to an append-only, hash-chained audit log, through a dedicated insert-only database role that cannot update or delete existing entries.

The platform is fail-closed by design: if a security-relevant dependency is unavailable, it refuses the request rather than serving data it cannot properly authorise.

ceadr engages the following sub-processors to deliver the service. Each receives only the data it needs for its stated purpose. We give 30 days notice before adding a new one; to object, email julian@ceadr.ai.

• Anthropic

  policy

  Claude API for scoring, contradiction detection, ontology, summaries. API inputs are not used to train models.

  United States

• AssemblyAI

  policy

  Audio transcription with speaker diarization and chapter summaries.

  United States

• OpenAI

  policy

  Optional fallback path for audio transcription via Whisper. The MCP deployment helper also references OpenAI as one of the supported agent endpoints.

  United States

• Voyage AI

  policy

  Embedding model for document chunks (semantic search and contradiction pairing).

  United States

• Nango

  policy

  OAuth brokerage and proxied API access to your connected SaaS providers (Google, Notion, Atlassian, Slack, GitHub, Jira).

  United States

• Clerk

  policy

  Authentication and session management. Stores user identifier, email, and session metadata.

  United States

• Resend

  policy

  Transactional email (workspace invites, agent notifications, writeback confirmations).

  United States

• PostHog

  policy

  Product analytics and pageview capture. Only loaded when the visitor opts in via the cookie banner.

  United States

• Sentry

  policy

  Error and performance reporting. When our servers hit an error, and for a sample of the requests they handle without one, they send Sentry technical details of the request, which can include the page address, your browser and operating system, your workspace, and the names of files involved. These details can identify you. In your browser, Sentry runs only if you opt in through the cookie banner.

  European Union (Germany)

• Google (Calendar Appointments)

  policy

  Booking links for the demo / discovery call flow on the marketing site. Calendar bookings are scheduled directly with Google.

  United States, multi-region edge

• Vercel

  policy

  Hosting for the ceadr.ai web application. Receives all frontend traffic and edge logs.

  United States, multi-region edge

• Railway

  policy

  Hosting for the ceadr API engine and worker processes.

  United States

• Supabase

  policy

  Managed Postgres database and file storage. Holds your stored data, including document text split into chunks with vector embeddings, the knowledge graph, analysis results, agent queries and answers, workspace and member records, the audit log, uploaded files, copies of some original source files, data export bundles, and contact details submitted on our website.

  European Union (Stockholm, Sweden)

We are pre-certification, and we say so plainly: ceadr is not yet SOC 2 or ISO 27001 certified. We would rather tell you that than imply a badge we have not earned.

Today we build to the same controls those audits check: tenant isolation, encryption, least privilege access, tamper-evident audit logging, and documented incident response and disaster recovery.

Next, as customers need them, we pursue formal third-party attestation and independent penetration testing.

Procurement needs specific evidence? Tell us what, and we will share our current posture and timeline.

We welcome responsible disclosure. If you believe you have found a security issue, email julian@ceadr.ai with "Security" in the subject line. Please give us a reasonable window to investigate and remediate before any public disclosure.