- 1Heath listens to the consultation, drafts the note and prescription, and verifies every line against what was actually said.
- The doctor confirms only what is uncertain, then slides to sign.
- We compared rules alone, a model alone, and both together. Only both together was useful and safe[1].
- Nothing is proven on real voices yet; that is the next step.
Design Principles
- 1
Zero extra work for the doctor. The doctor talks as usual; the record writes itself.
- 2
The doctor decides. Nothing is signed by the software. Anything uncertain is shown for the doctor to confirm.
- 3
Nothing invented, nothing lost. The system never fills in a drug, strength or dose that was not said. Anything it is unsure of is flagged, never dropped.
- 4
Indian reality first. Code-mixed speech, brand-name prescribing, two-minute visits, ceiling fans and corridor noise, AYUSH alongside modern medicine.
- 5
One record, many views. Each person in the visit gets their own app on the same consented record.
- 6
No side between systems. 1Heath records what patients take from every system and never tells anyone to stop or start one.
How It Answers Each Root Cause
| Root cause | 1Heath's answer |
|---|---|
| No time to take a history | The history is captured from the conversation; nobody types |
| Paper does not travel | A signed digital prescription and note, on the patient's phone and in the record |
| Care split across providers | One record per patient, shared with consent, linked to the ABHA health ID |
| AYUSH medicines never mentioned | An "Also taking" list: the desk asks and the scribe listens for medicines from every system |
| No history in an emergency | An emergency card on the patient's lock screen and as a QR code |
| Inherited disease unseen | With consent, family records link, so one diagnosis shows which relatives should be tested |
One Visit, Step by Step
- DESKRegister and consentPatient found or registered, ABHA linked, recording consent, vitals
- DOCTORTalkSwipe to start; the doctor and patient talk as usual
- 1HEATHDraft and verifyNote and prescription drafted, every line checked
- DOCTORConfirm and signResolve the few flagged lines, slide to sign
- EVERYONESharePharmacy, patient and family see their part of the record
How a Line Earns Trust
Every prescription line ends in one of three states. The doctor sees the doctor's own words next to what the system understood.
| State | Meaning | Doctor's action |
|---|---|---|
| Verified | Every part matches what was said and the catalogue | Nothing |
| Confirm | A name, strength, dose, frequency or duration is uncertain or missing | One tap on a choice; none is pre-selected |
| Cannot Sign | Conflicts with a recorded allergy, or a banned combination | Change the line |
The checks cover allergies, pregnancy, children's doses by weight, duplicates (including medicines the patient already takes), weekly drugs written as daily, banned combinations and look-alike brand names. The full rule list is on the architecture page.
Alternatives We Compared
We did not start with this design. We tested the obvious options on the same unseen simulated consultations[1].
| Option | Why not, on its own |
|---|---|
| Doctor types the history | Adds typing to a visit that lasts about two minutes |
| Human scribes | A salary many small clinics cannot carry, and no shared record after the visit |
| A language model alone | Captures the most, but adds facts nobody said. One invented allergy or dose is unacceptable |
| Rules alone | Never invents, but misses most of the conversation and misread who refused what |
| Global scribes | Built for English, long visits and insurance billing, not code-mixed speech, brand names or AYUSH |
| Model plus rules (1Heath) | The model understands the sentence; the rules decide whether each line can be trusted; the doctor resolves the rest |
Why We Think It Is the Best Available Design, and What Could Prove Us Wrong
The case for it. Missing medicines are the most common history error[2], so the record must be written at every visit without effort. Invented clinical facts are the biggest risk of language models, so every line must be checked against the words spoken. Uncertainty must go to the doctor as a choice, because a guess that looks confident is worse than a question. No single component does all three; the combination does.
What is still unproven. Every result so far uses synthetic voices[1]. Real clinics are noisier, patients interrupt, and accents vary. Other Indian scribes are already in daily use[3], and India still has no shared, real-world benchmark for comparing them[4]. Until 1Heath is measured on consented real consultations, "best" is a design argument, not a measured fact.
Beyond the Consultation
Emergency Card
Blood group, allergies, long-term conditions, current medicines from every system and two family contacts, on the phone's lock screen and as a QR code, readable without unlocking. The patient chooses what it shows.
Family Health
With each person's consent, a family's records link. When one member is diagnosed with an inherited condition such as thalassaemia trait, the doctor sees which relatives should be offered testing, with counselling first.
Watch Data
Patients who connect a smartwatch can share sleep, steps, active minutes and resting heart rate as simple graphs, computed the same way for every brand, shown as context and never as a diagnosis.
What 1Heath Is Not
- Not a diagnosis tool: the doctor diagnoses; 1Heath records and checks.
- Not a replacement for clinical judgement: every prescription is the doctor's, signed by the doctor.
- Not a data business: patient data is never sold.
- Not ready for real patients yet: see what must happen first.
Sources
- 1Heath internal benchmarks, October to November 2026. Synthetic consultations; method described on the Models and data page. pawankalyan.dev
- Tam VC, Knowles SR, Cornish PL, et al. Frequency, type and clinical importance of medication history errors at admission to hospital: a systematic review. CMAJ, 2005;173(5):510. cmaj.ca
- Sarvam AI. How HealthPlix turns doctor consultations into medical records with Sarvam, 4 June 2026. sarvam.ai
- Jaiswal SD, Krithi S, Makani A, Datta S, Sitaram S, Jain M. Evaluating ambient clinical scribes in India: the need for multilingual real-world clinical conversation data. arXiv:2609.17355, September 2026. arxiv.org