How does on-device visit capture become a structured note a clinician can sign off?

This page describes a planned PatientSaathi method, not a live product. You cannot record a visit here, and you cannot get a signed note from this site. Ask Sathi on the homepage is a demo. It does not listen to a consult and it does not write a note.

The product we are describing is a three-step pipeline: capture on the device, turn that capture into a structured draft with mismatch flags, then stop until a licensed clinician signs off. It is not a chatbot, not an “AI doctor”, and not a diagnosis.

What this is, and what it is not

  • Not live. PatientSaathi does not store visit audio, does not store family records, and does not book doctors on this site.
  • Not a diagnosis. A structured draft is admin/text organisation. It does not decide what is wrong with a patient and it does not prescribe.
  • Not a silent write. Nothing reaches a chart, locker, or shared record unless a licensed clinician reviews and signs it.
  • No self-improving memory. The pipeline does not keep a hidden model that trains on your visits.

If you need care today, talk to a licensed clinician. In a medical emergency in India, contact 112.

Step 1 — On-device visit capture

Audio stays on the device that recorded it. There is no bot that joins the consult as a participant.

The UX analog is HyNote (on-device meeting speech-to-text, no bot in the room). HyNote for Mac appeared on Product Hunt around 20 August 2026 as a local-transcription pattern: system/device audio, transcription on the machine, no extra attendee in the call. That is a pattern citation only. HyNote is not a PatientSaathi partner, vendor, or endorsement, and we are not claiming their accuracy, user counts, or medical use.

What we mean by this step, if we build it:

  1. The clinician (or the clinic device) starts capture on that device.
  2. Speech-to-text runs locally. The audio file is not uploaded to PatientSaathi as a side effect of capture.
  3. The raw transcript is a draft input, not a medical record, until someone signs.

We will not claim we already run this in Delhi clinics.

Step 2 — Structured note (sections + mismatch flags)

The raw transcript is messy. The next step is to put sentences into named sections and flag internal mismatches so a human can see them. This is organisation and QA of text, not a clinical opinion.

Research method we are borrowing from, not validating. Hartsock, Lam, Otteni, Qayyum, Gatenby, Araujo, and Rasool describe a locally deployed four-agent pipeline for radiology-report structuring and quality assurance (arXiv:2608.18072, 19 August 2026). Their retrospective set was 638 CT reports of the chest, abdomen, and pelvis, dictated with voice recognition. They used three agents to place Findings sentences into anatomical sections (rules, then local LLMs) and a fourth agent to flag mismatches (including Findings vs Impression, polarity/laterality issues, gender–anatomy conflicts, and undocumented communication of critical findings). The authors note voice-recognition / speech-to-text error as a reason QA is needed. Two radiologists independently reviewed a 45-report subset; both agreed that 31 reports (69%) were correctly restructured.

PatientSaathi has not run that evaluation. We have not processed 638 reports, we have not measured 69%, and we do not claim their QA scores. Their work is radiology reports behind a hospital firewall. A clinic visit note is a different document. We cite the paper as a method shape: local pipeline, sections, mismatch flags, human review of flags.

What a PatientSaathi draft would try to do, if we build it:

  • Keep the speaker’s words. Do not invent findings that were not said.
  • Put text into labelled sections a clinician already uses (for example chief concern, history, exam, plan) — the exact headings are not live yet.
  • Flag mismatches for a human (laterality, negation, section vs summary). A flag is a question, not a diagnosis.

Step 3 — Human sign-off required

The draft stops. A licensed clinician has to read it, change it, and sign it. Until that happens:

  • It is not a signed note.
  • It is not written to a hospital file, ABHA locker, or PatientSaathi locker (we do not have a live locker).
  • It is not sent to another clinic.

No silent writes. No auto-file. No model that “learns” a patient over visits in the background.

What PatientSaathi can and cannot do today

  • Cannot: record your visit, store the audio, produce a signed note, book the clinician, or check symptoms.
  • Can: publish this method page, and take an interest registration if you want a callback when tools like this are actually available.
  • Ask Sathi is a demo. It does not capture a visit.

FAQ

Is this an AI doctor?
No. It is a planned capture-to-draft-to-sign-off method. The clinician signs. The software does not.

Does audio leave the device?
The design we are describing keeps capture on-device. That is not live on this site. Do not send visit recordings to PatientSaathi through the interest form.

Did PatientSaathi test this on 638 reports?
No. The 638-report / 69% figures are from Hartsock et al., arXiv:2608.18072, on radiology CT reports. We have not repeated that study.

Is HyNote part of PatientSaathi?
No. It is a public UX analog for bot-free, on-device transcription.

Do I need this to visit a clinic in Delhi?
No. A normal walk-in still uses the clinic’s own file. See What should I carry to a first clinic visit in Delhi?.

Not a doctor. This page is a method description for a planned PatientSaathi pipeline. It is not a diagnosis, a prescription, a care plan, or a live scribe. PatientSaathi does not store visit audio and does not book doctors on this staging site. If you feel unwell, talk to a licensed doctor. For a medical emergency in India, contact 112.

Scroll to Top