Case narratives, drafted in minutes. Reviewed and approved by you.
Nara AI turns a structured adverse-event case into a clear, chronological draft narrative, so your safety team spends its time reviewing rather than typing. Nothing is final until a qualified reviewer signs it off.
A 54-year-old female patient experienced Rash and Pruritus while receiving Examplezide for hypertension. Relevant medical history was not reported. The patient was treated with Examplezide 10 mg, administered orally, once daily from 03 March 2026; Examplezide was withdrawn on 16 March 2026. Both events began on 15 March 2026 and had resolved by 22 March 2026; the outcome was recovered/resolved. The case was non-serious. The reporter assessed the causal relationship between Examplezide and both events as possible.
How it works
Three steps. You keep the pen.
No installation, no change to your safety database. Nara AI works alongside the tools you already use.
Share the case
Upload the case export you already have, or type the facts into a short form: the patient's age and sex, the medicines, the reactions and what happened when.
Nara AI drafts and checks
In seconds you get a readable, chronological narrative in the standard format. Automatic checks then compare the draft with the case, line by line.
You review and sign off
Read it, edit it, approve it. Your sign-off is recorded with a fingerprint of the exact text you approved, so the record is yours and tamper-evident.
Why safety teams can trust it
Designed to assist, never to decide.
A drafting aid for regulated work has to be conservative. These four principles are built into the product, not written into a policy.
You stay in control
Every narrative is a draft until a qualified person approves it. The tool cannot submit, send or file anything.
Nothing is stored
Case text is processed in memory and gone when the draft is done. What remains is a record of who did what, and when.
Stays in Europe
The application, its records and the AI processing all run inside the EU, under European data-processing terms.
Checked before you see it
Automatic checks compare each draft with the case. Anything off is shown to you, not hidden.
For QPPVs, PV specialists and IT reviewers
The specifics.
Everything above, in the terms you would use in an assessment.
What it produces
- A draft of the E2B(R3) H.1 case narrative in the chronological convention, from structured, already MedDRA-coded case data.
- MedDRA fidelity: every reaction term appears verbatim (PT/LLT) as supplied. Nara AI is not a coding engine and never re-codes.
- Seriousness, causality and expectedness are never inferred, upgraded or downgraded by the tool: the prompt forbids it, and a draft that DENIES a value the case supplies is flagged for the reviewer. What is not machine-verified is the wording of a free-text causality assessment against the value supplied - that comparison is the qualified reviewer's, and we say so rather than imply a check we do not run.
- Input: E2B(R3) XML, CIOMS-style CSV or a web form. Output: text, copy or download (.txt/.docx). No E2B XML is generated.
How it is controlled
- 14 deterministic gates run after every generation: MedDRA fidelity, no fabrication, structural completeness, EU personal-data masking (GVP Module VI Addendum II) and cross-field consistency (dates, doses, outcomes, seriousness language, action taken, death completeness). Two of the fourteen check a draft against an expected value carried only by our own reference cases, so they do not apply to a case you upload; each draft reports how many gates were applicable rather than assuming all of them were.
- Draft-only by design. Status is tracked by the system, not embedded in the text; the approved narrative is clean. No code path transmits to EudraVigilance, FAERS or any gateway, enforced by tooling.
- Human sign-off is the control. The approval binds a SHA-256 digest of the approved text in an append-only audit trail. Validation-ready (GAMP 5, Category 5 context), not validated.
Where the data goes
- Zero-persistence: narrative and case content are never written to disk. The audit trail holds metadata only: user, case reference, timestamps, content hashes, gate results, reviewer notes.
- EU-resident end to end: application and audit database on Render in Frankfurt; model inference on Amazon Bedrock, sent to eu-north-1 through an EU-only cross-region inference profile, so a request may be served by another EU region and never leaves the EU, under the AWS Data Processing Addendum, which does not store prompts or outputs.
- Direct identifiers are removed deterministically before any model call. Participants enter synthetic or de-identified cases only.
Access and records
- Sign-in via Auth0 (e-mail and password), invite-only, one workspace per organisation, roles (reviewer/admin).
- Every generation, edit, sign-off and consent is an audit event; participants can review their own trail in the app.
- Leaving is one click; account deletion is a two-step request, both recorded. See the Privacy Policy and Terms.
The evaluation pilot
Try it on your own cases, for free.
A small number of pharmacovigilance consultants and safety teams are invited to use Nara AI for a fixed period and tell us, plainly, where the drafts fall short. You get the tool and a direct line to the person who builds it; we get the feedback that shapes it.
- Free and invite-onlyNo card, no contract to sign beyond the short participation terms.
- Your cases, de-identifiedBring the kind of cases you actually write; remove names and identifiers first.
- Honest feedback wantedA short note per draft is enough. We would rather hear what is wrong than what is nice.
- Leave any timeOne click in the account menu. No retention calls, no forms.
Questions
Frequently asked
Do I need to install anything?
No. Nara AI runs in your browser. You sign in, share a case, and review the draft. Nothing is installed and nothing connects to your safety database.
Is the output ready to submit?
No, and it is never meant to be. Every narrative is a draft until a qualified reviewer edits and approves it. The approved text is yours to use in your own process; Nara AI itself has no way to submit anything.
Is it validated?
It is validation-ready, not validated. The design follows GAMP 5 thinking for a Category 5 drafting aid where the human sign-off is the control. Evidence such as the golden-set regression record and the deterministic gate suite is available to pilot participants on request.
Which language does it write in?
English, in the chronological convention used for E2B(R3) H.1 narratives.
What data do you keep?
Not the case and not the narrative. We keep an append-only record of activity: your e-mail, the case reference you entered, timestamps, content fingerprints (SHA-256), check results and your notes. The Privacy Policy lists everything.
Where is my data processed?
Inside the EU: the application and its records run in Frankfurt, and the AI model runs on Amazon Bedrock in EU regions only, under AWS's data-processing terms. Nothing is sent outside the EU.
Can I use real cases?
Only de-identified ones, and only cases you are entitled to use. Remove names, initials, dates of birth, record numbers and real identifiers first. If a case belongs to a client, check with them before you upload it.
Who is behind Nara AI?
An independent founder, building it with pharmacovigilance practitioners during this pilot. Write to karl@naramed.ch; you will get a reply from the person who builds it.