Skip to content

Tara AI

Privacy policy

Tara AI is a browser extension for authorised staff of the International Medical Affairs Office (IMAO) at Hamad Medical Corporation, Doha, Qatar. She runs in Chrome or Edge. Nothing is installed on the PC. This page describes the store package: it does not replace the browser's new-tab page.

Effective date: 16 October 2026

Contact: drkhalifaalqaatri@gmail.com

Questions about this policy go to the contact above, or to the support page. The IMAO portal at imao.hamad.qa is an intranet site and is not a public contact address.

Where the data is hosted

The Tara Supabase project (https://gqisjpbeuzklawkaoege.supabase.co) is hosted by Supabase on Amazon Web Services in region ap-south-1 (Mumbai, India). People using Tara are in Qatar. Account data, portal content, patient information, telemetry, and files kept in that project are a cross-border transfer from Qatar to India. Row-level security is the access boundary.

What she handles

  • Account. Sign-in uses an @hamad.qa email and a one-time code. The session is stored by the Tara Supabase project and in the browser.
  • Portal and health information. Medical-travel requests, committee notes, and clinical documents the person opens on the IMAO portal are stored in chrome.storage.local and in the Tara Supabase project. Patient data does leave the browser for that project. Case and patient records are not deleted on a schedule.
  • Microsoft Graph profile. /me returns the staff member's name, job title, and photo. Tara caches that profile in the browser. The side panel Email tab uses the name when composing a message. The same profile is shown as a greeting only on Tara's own page, and only if someone opens that page by its extension URL. The store package does not make that page the browser's new tab.
  • Microsoft Graph calendar. /me/calendarView supplies the upcoming-events list on that same Tara page. It is not a new-tab override.
  • Mail. The Email tab and the portal mail chip read and send messages through the person's Microsoft 365 mailbox (/me/messages, /me/sendMail, and Outlook on the web). Those surfaces are in the store package. The full message body stays in the mailbox. When the person classifies patient mail, Tara sends the subject, body preview, and sender to a Supabase function, which calls Gemini. A case digest or a question about that mail also sends the patient name, QID, and HC number. A question about a document also sends the extracted document text.
  • Authentication secrets. The Microsoft Graph token is harvested from a Graph Explorer session the person has already consented to. MRTA and NAS credentials are filled into those sign-in pages. Credentials are not sent to Gemini or Deepgram. The store build does not tick the NAS reCAPTCHA checkbox and never attempts the image challenge.

Who receives data

  • Tara Supabase project, hosted in AWS ap-south-1 as described above, and the storage host https://gqisjpbeuzklawkaoege.storage.supabase.co. Authentication, database rows (including patient and portal content), the Storage buckets listed below, and Realtime.
  • Google Gemini is a processor. Supabase Edge Functions send the document text, portal excerpts, or dictation phrases the person submitted so Gemini can perform that AI feature. Patient-mail classification sends the subject, body preview, and sender. A digest or a question also sends the patient name, QID, and HC number, and a document question sends the extracted text. The full message body stays in the mailbox. Gemini does not supply Google user data to Tara. The extension does not open a Google AI host. The model that runs is chosen on the server. A usage row records the model name, token counts, the staff member's email prefix, the function, latency, and an error kind. It does not record the prompt or the document.
  • Deepgram at wss://api.deepgram.com, a direct connection from the browser after an edge function mints a short-lived token. Microphone audio for live dictation goes there.
  • Microsoft Graph and Outlook on the web. Profile, calendar, and mail, as described above.
  • Google Fonts. On IMAO portal pages Tara injects an Inter stylesheet from https://fonts.googleapis.com. The request is a font CSS file. The browser may send the portal URL, which can include a request id, as the Referer.
  • The IMAO portal, MRTA, and NAS (/idp/*only). These are the pages the person is already using.

Storage buckets

The extension uses these Supabase Storage buckets:

  • portal-captures (private) — a portal-capture archive: DOM, styles, MHTML, and the screenshots packed into that archive. These objects are not deleted on a schedule.
  • mobile-files (private) — files the mobile bridge exchanges. Objects older than 1 day are deleted.
  • facility-logos (public) — hospital logos the portal pages display.
  • country-flags (public) — flag images looked up from the countries table.
  • releases (public) — an office install downloads a verified ZIP from here. This store package does not.

Bug-report screenshots are not stored in a bucket. Up to four images are saved as data URLs in the bug_reports table. The debug_reports table holds a redacted diagnostic report, not screenshots. There is no debug bucket. The extension does not use the uploads, intake, ta-attachments, or assets buckets.

Co-browse

Co-browse is remote control of a portal tab, not a view-only stream. Frames are checked against the IMAO portal or a page of this extension, then broadcast on Supabase Realtime. The operator's clicks and keys are delivered back and replayed on the page. The frames leave the machine. They are live broadcasts. They are not stored as a video archive. A session that stops heartbeating is marked expired after 2 minutes. Ended or expired session rows are deleted after 1 day. The audit record — who operated, the event, the portal hostname and page name, and an action count — is deleted after 30 days.

Telemetry

Presence rows record the user id, email, extension version, when the person came online, when they were last active, whether the portal session is signed in, how many tasks are in flight, and whether the shared-folder and travel workers are online. There is no update-mode field. Presence rows are not deleted on a schedule.

The project also stores automation run rows, feature-usage rows, and ai_usage_events (model name, token counts, the staff member's email prefix, the function, latency, and an error kind — never the prompt or the document). client_errors stores an error name, message, stack, page path, and install id. client_health_events stores a kind, a severity, a payload, and an install id. Both error tables can include the signed-in email. This is operational telemetry. It is not sold and it is not an advertising SDK. Patient-identifying information is not supposed to be written into these rows, into logs, or into thrown errors.

Tara does not store a browsing history. She may read other tabs' URLs in memory so co-browse can find its own page, an automation can reuse a portal tab, or a temporary bypass can see the tab it was asked about. Those URLs are not saved as history.

Chrome Web Store User Data Policy — Limited Use

Tara's use of user data follows the Chrome Web Store User Data Policy, including the Limited Use requirements.

She uses that data only to provide the features the person is using:

  • Email — reading and sending mail from the side panel Email tab and the portal mail chip.
  • Profile — the person's name, job title, and photo, for the Email tab and, if they open Tara's own page by its extension URL, for the greeting on that page.
  • Agenda — upcoming events on that same Tara page. The store package does not replace the browser's new tab.
  • An AI feature the person invoked — the document, phrase, or mail excerpt they submitted, processed for that request by Gemini acting as a processor.

She does not sell this data. She does not use it for advertising or for serving ads. She does not allow it to be used for determining creditworthiness or for lending purposes. She does not transfer it to advertising platforms. Microsoft Graph profile and calendar are sent to Microsoft for the Graph calls the person asked for. The profile cache stays in the browser. Mail is read and sent through Microsoft Graph and Outlook. When the person classifies patient mail, the subject, body preview, and sender go to a Tara Supabase function, which calls Gemini. A case digest or a question about that mail also sends the patient name, QID, and HC number. A question about a document also sends the extracted document text. The full message body stays in the mailbox. Dictation phrases and other document text submitted for an AI feature go to the same project, whose functions call Gemini as a processor.

Retention

Scheduled deletion in the Tara Supabase project:

  • The Microsoft Graph access token lasts about 60–90 minutes. There is no refresh token. It is cached in the browser until it expires. The Graph profile and photo cache in the browser is refreshed about every 12 hours.
  • Co-browse sessions are marked expired after 2 minutes without a heartbeat, and those session rows are deleted after 1 day. Co-browse audit rows are deleted after 30 days. Frames are not kept as a recording.
  • client_errors and client_health_events are deleted after 45 days.
  • ai_usage_events are deleted after 400 days.
  • Automation run rows (rpa_execution_results) are deleted after 400 days. Clean completed steps are deleted after 30 days. Steps kept as incident evidence stay until that run is deleted.
  • Workflow-event rows are deleted after 400 days. The travel-timeline access log is deleted after 400 days. The sealed travel-timeline cache follows the office retention setting (30 to 730 days, default 365). A source that has been missing longer than the missing-source grace (7 to 180 days, default 60) is dropped. Extractions left without a source are dropped after the orphan grace (1 to 90 days, default 30). Quota rows are deleted after 7 days.
  • Team chat messages are deleted after 90 days. Office digest snapshots are deleted after 90 days. Unpinned desktop conversations that have not been archived are deleted after 30 days. An archived desktop conversation is kept. Unpinned mobile conversations are deleted after 30 days. Assistant requests that have finished (done, error, or cancelled) are deleted after 7 days. A request that never finishes is kept. mobile-files objects are deleted after 1 day. Finished reminders are deleted after 30 days.
  • Case-command rows are deleted after 30 days. Document-intake files are deleted retain_days after that request closes, and intake event rows are deleted after 30 days. An import batch is purged on the date set for that batch.

These have no automatic deletion: presence rows, portal-capture archives, case and patient records, job-list history, bug reports, and debug reports. Row-level security still limits who can read them.

What she does not do

Data is not sold. There are no ads and no advertising identifiers. The store package does not offer a ZIP download of a newer build, does not replace the new-tab page, and does not tick reCAPTCHA.

Support