CLARA · Security & Trust

EU-hosted, separated per company, on the record.

What is implemented today and checkable in the product. Legal pages (Impressum, privacy policy, terms) follow the legal review. Until then, request our DPA at hello@odradekai.com.

Hosting & data residency

  • The CLARA API and application data run in the EU: a dedicated server in Germany (Hetzner) and an EU-region Postgres database (Supabase).
  • AI runs on EU-hosted models (Mistral). EU-only mode is on by default and refuses model providers outside the EU.
  • The website and dashboard front end are served through Vercel's CDN. Customer feedback is processed by the EU-hosted API, not the CDN.
  • Self-hosting is supported: the API is a standard Docker deployment and can use EU-hosted or fully local models.

Tenant isolation & access

  • Each company's data is separated by row-level security in Postgres. The API connects through a database role that cannot bypass it and refuses to start through one that can.
  • Identity and tenant come only from the verified login token or an API key. Access is enforced on the server, by role.
  • Two-factor sign-in (TOTP) is available, and repeated sign-in attempts are rate-limited per IP.
  • Only the operator can change the AI endpoint. A workspace admin cannot redirect feedback to another provider.

Security measures

  • TLS everywhere with HSTS, plus strict security headers on every web and API response.
  • Webhooks need an HMAC-SHA256 signature over a timestamp and the body, valid for five minutes, so a captured request cannot be replayed.
  • Connector URLs pointing at private or internal addresses are refused. API keys are stored as hashes and shown once.
  • Audit records are append-only. Evidence packs carry a content hash, so what was approved stays verifiable.

Data subject rights & governance

  • GDPR Art. 17 (erasure) and Art. 20 (export) are product endpoints, not ticket queues.
  • Common identifiers (emails, phone numbers, IP addresses, customer and account IDs) are removed from feedback before it reaches a model. Reviewer names from public review sites are pseudonymised on arrival.
  • No action is pushed to Jira or Slack without a human approval. With four-eyes on, two different people approve, and exactly the approved version is sent.
  • Works-council mode (§87 BetrVG): below admin, person-level fields become role labels. No per-employee metric can be derived.
  • EU AI Act Art. 50: outbound text that no person reviewed carries an AI-disclosure line, and human-approved content records its reviewer. Human review is not an exemption: machine-readable marking of AI-generated text (Art. 50(2)) is not yet applied to text pushed to Jira or Slack, and remains open.

AI transparency

  • Provider-agnostic: any OpenAI-compatible endpoint, including EU-hosted and self-hosted models. No fine-tuning on customer data.
  • A live model card shows the configured model and the published accuracy per language, with sample sizes, dataset date and the model it was measured on, and says when the configured model has no published figures.
  • Every output carries a model score (labelled as the heuristic it is), evidence links and stated limits. Thin evidence gets a refusal, not a guess.
  • Outcome readouts are graded A to E by measurement design. Manual entries are labelled as such and never mixed with instrumented measurements.

Certifications & roadmap

  • ISO 27001: on the roadmap. We pursue it as revenue and data volume grow.
  • BSI C5 (Type 1, then Type 2): planned after ISO 27001. It doubles as NIS2 and DORA supply-chain evidence.
  • DORA ICT annex and NIS2 supply-chain documentation are prepared for regulated pilots. A signed DPA (AVV) is available on request.

Questions, security reports or DPA requests: hello@odradekai.com

Security & Trust · CLARA