Privacy Policy - Nara AI (Evaluation Pilot)

Last updated: 7 September 2026. This is the evaluation-pilot version. Where it conflicts with the signed participation agreement, the participation agreement prevails.

Nara AI ("the Service") is an evaluation-pilot drafting aid for pharmacovigilance case narratives, built and operated by its founder as a sole proprietorship ("we", "us", "the operator"). This policy explains what personal data we process when you use the Service during the pilot.

Who we are

Operator and data controller: Nara AI - Karl Bast, sole proprietorship, Huobweid 5, 6333 Hünenberg See, Switzerland. Contact for any privacy question or request: karl@naramed.ch.

What personal data we process

• Your account and sign-in. Sign-in is handled by our identity provider, Auth0 (an Okta company), which processes your email address, password hash, and authentication metadata. Access is invite-only: we create your login after you have been accepted into the pilot. No social sign-in (Google or other) is offered.
• Your activity (audit trail). We keep an append-only audit trail of metadata about your use: your email address, the case reference you enter, timestamps, content hashes (SHA-256) of each draft and of any narrative you sign off, validation-gate results, and any feedback comments or reviewer notes you submit. This is a regulatory-style record; it is immutable and cannot be edited or deleted by you.

Our legal basis

• Your account and sign-in: performance of the participation agreement with you (Art. 6(1)(b) GDPR).
• The audit trail: our legitimate interest in keeping an integrity record of what the Service produced and who approved it, which is what makes the output defensible to you and to an auditor (Art. 6(1)(f)).
• Security, access control and abuse prevention: our legitimate interest in keeping the Service and its participants safe (Art. 6(1)(f)).
• The case content you submit: we do not decide why that is processed; you do. For it we act as your processor, on your instructions and on your legal basis. See "Two roles" below.

Two roles

Two different things happen in this Service and we hold a different role in each. Reading them as one is the usual source of confusion.

• Your account and the audit trail: we are the controller. This policy describes that processing.
• The case data you submit: you are the controller (or your client is, and you are their processor); we are your processor and act only on your instructions. That relationship is governed by the Data Processing Agreement (Annex 1 to the participation terms), which is what Art. 28 GDPR requires and which we provide before you are given access.

A case report stripped of names and initials is pseudonymised, not anonymised - it remains personal data, and it is data concerning health. We treat everything you submit that way.

What we do NOT store

We do not store narrative text or case content. The de-identified case you submit and the model's draft output are processed in memory to produce the draft and are not retained on our servers (zero-persistence). Only the metadata and hashes listed above are recorded.

The case-data rule

The Service is for synthetic or fully de-identified cases only. You must not enter patient names, initials, dates of birth, medical record numbers, addresses, or real case identifiers. The built-in identifier masking is a safeguard, not permission; you remain responsible for de-identifying your inputs before submitting them.

Where your data goes (sub-processors and transfers)

• Auth0 (Okta): sign-in and identity.
• Render: application hosting and the audit database, located in Frankfurt (EU).
• Amazon Web Services (AWS): the AI model (Anthropic Claude, operated by AWS on Amazon Bedrock) that drafts the narrative. The de-identified prompt and the model's output are processed inside the EU: the request is sent to eu-north-1 (Stockholm) through an EU-only cross-region inference profile, so it may be served by another EU region, and it never leaves the EU. Processing is under the AWS Data Processing Addendum. AWS operates Bedrock on a zero-data-retention and zero-operator-access basis: by default it does not store model inputs or outputs, and no AWS operator can access them. A small number of models are excepted and do require retention of up to 30 days for automated abuse detection; the model this Service uses is not one of them, and we will not move to a model that requires retention without updating this policy first. Where AWS does retain data for those excepted models, it stays within AWS and is not shared with the model vendor. Nothing you submit is used to train, fine-tune or improve any model.

The de-identified case therefore stays in the EU end to end (hosting and audit database in Frankfurt, model inference in Stockholm). Should the operator ever route inference outside the EU, this policy and the participation agreement will be updated first and the transfer safeguards set out there.

How long we keep it

Account and audit-trail records are kept for the duration of the pilot and for 12 months afterwards. At that point your workspace's records are exported to you, if you want them, and deleted. You can ask us to delete them sooner and we will, unless we need a record to establish, exercise or defend a legal claim (Art. 17(3)(e)). Case content and draft narratives are not kept at all - see "What we do NOT store".

Your rights

Subject to applicable law you may ask us for access to the personal data we hold about you, for its correction, for its erasure, for restriction of its processing, for a copy in a portable format, and you may object to any processing we base on our legitimate interests. You may also complain to your data-protection supervisory authority.

Two honest qualifications. The audit trail is append-only by design - that is what makes it trustworthy as a record - so entries are not removed one at a time on request; erasure is handled by deleting your workspace's records through a controlled operator process, which you can start yourself at any time from the account menu ("Delete account"). And we may keep a record where we need it to establish, exercise or defend a legal claim.

To exercise a right, contact karl@naramed.ch.

Security

Access to the Service requires authentication. The audit database is a paid, EU-hosted PostgreSQL instance running under an append-only, minimum-privilege role. We do not sell personal data.

Changes to this policy

We will update this policy as the pilot evolves and will change the date at the top when we do.

Contact

Nara AI - karl@naramed.ch.