Rxnote AI

Complete patient intelligence. Zero documentation burden.

Comment

Rxnote — The Clinical Memory Layer

Complete patient intelligence. Zero documentation burden. In every language the visit actually happens in.

Inspiration

A family member of ours repeated their full medical history to three different clinicians in one week. None of them had the whole picture. That is not an edge case across India, Nepal, and Southeast Asia — that is the default state of care for 2.16 billion people.

What struck us is that both sides of the exam room lose the same conversation, at the same moment, for different reasons.

The clinician loses it to documentation. Roughly 40% of session time goes to notes — written by hand, after hours, time stolen directly from patient care. Discharge summaries take 4–10 days to assemble from scattered inputs. Every consultation starts from scratch, with no carry-forward context from the last one.

The patient loses it to memory. Studies show patients forget up to 80% of what a doctor tells them within minutes of leaving the room, and that missed follow-through drives billions in avoidable cost.

So the most information-dense event in all of healthcare — the consultation itself — evaporates the second it ends. Twice.

And every product built to fix this quietly assumes a single-language, English-speaking visit. That is not how healthcare happens here. A real consult in Chennai or Kathmandu moves between Tamil, Hindi, Nepali, Telugu, Malayalam, Kannada and English — sometimes inside a single sentence. Existing tools don't just perform worse on these visits; they break, silently dropping or garbling the non-English segments.

PatientDrive exists because the conversation should outlive the appointment — for the doctor and the patient — and because "never forget what your doctor said" shouldn't only be true if you happen to speak English.

What it does

One recording. Two memories. A layer that compounds.

PatientDrive captures a consultation — in-person or tele, in any mix of languages — and generates two fundamentally different artifacts from the same audio, for the two people who need it most.

For the clinician — zero documentation burden

  • Ambient AI scribe. Structured SOAP notes generated instantly across 40+ specialties. Our longest recorded consultation ran 1 hour 41 minutes and was documented in under 5 minutes.
  • 10+ Indian languages with native code-switching. Tamil, Hindi, Telugu, Malayalam, Kannada, Nepali, English — mid-sentence switches preserved, not lost.
  • Discharge summaries at a click. Structured, accurate, file-ready — collapsing a 4–10 day assembly process into a single action.
  • Secure audio archive. Every session stored immutably for medico-legal reference, DPDP Act 2023 compliant.
  • Clinical task extraction (in dev). Spoken decisions become tracked tasks — lab orders, referrals, WhatsApp follow-ups — instead of dying in the note.
  • Medical record analyser (beta). Analyses historical records and surfaces relevant history, risks, and gaps before the next visit, so no round starts blind.

For the patient — complete comprehension

  • Record the visit. Start recording and stay present, instead of scrambling to take notes.
  • Get a plain-language summary with next steps. The same consultation, rewritten for a patient rather than a clinician: what was discussed, what to do, which medications, which follow-ups — in their language.
  • Share with the care circle. A spouse, an adult child abroad, a caregiver — everyone stays aligned without the patient re-explaining the visit from memory.
  • Prep for the next visit. Patients jot questions ahead of time and get reminded before the appointment, so nothing gets missed.

The layer itself

This is the part that isn't just two features stapled together. Because both outputs come from the same captured session, memory accumulates. The clinician's next consultation opens with carry-forward context from the last one. The patient's next visit opens with the questions they wrote down and the plan they were given. The record analyser reads across the whole history, not one document.

The patient stops being the lossy transport layer for their own medical history. That's the clinical memory layer.

How we built it

Gemini as the reasoning core. The hardest problem in this product isn't transcription — it's that one messy, multilingual transcript has to become two artifacts with completely different registers, audiences, and failure modes. We use Gemini's long-context window to hold entire multi-hour consultations in a single pass and generate both the clinical SOAP note and the patient plain-language summary from one grounded read of the conversation, rather than summarizing a summary. Long context is also what makes carry-forward possible: prior visit notes are loaded alongside the current transcript so the model reasons across a patient's history, not just today's session.

Speech-to-text tuned for code-switching. We evaluated multiple engines on real South Asian doctor-patient audio and landed on Soniox for its handling of multilingual and code-switched speech — the single highest-leverage stack decision in the project, because everything downstream inherits transcript quality.

Mobile app in React Native. Native recording and summary experience on both iOS and Android — Android coverage being non-negotiable for patient reach across South Asia.

The full loop, end to end. Care-circle sharing, visit prep, reminders, discharge generation, and the secure DPDP-compliant audio archive — built out rather than stubbed, because the loop is what makes this a habit instead of a demo.

Challenges we ran into

Code-switched speech breaks most STT models. Getting clean transcripts when a doctor and patient move fluidly across two or three languages in one sentence required genuine evaluation work — most engines fail quietly, producing confident, fluent, wrong output for the non-English segments. Silent failure in a medical transcript is worse than loud failure.

Two audiences, one transcript, opposite constraints. The clinical note must preserve full clinical fidelity and terminology. The patient summary must strip all of it and stay comprehensible — without dropping anything that matters or drifting into advice the clinician didn't give. Getting both from one source, with no contradiction between them, took the most careful prompt design in the project.

Plain language without losing clinical meaning. The failure mode we were most afraid of wasn't a bad summary — it was a plausible one. We constrained generation tightly to what was actually said in the session, and designed the patient output to defer rather than fill gaps.

Building the whole loop inside a hackathon window. "Record → summary" demos easily. Sharing, reminders, discharge generation, and a compliant archive do not. We prioritized ruthlessly to ship a complete journey for both users rather than one flashy moment for one of them.

Accomplishments that we're proud of

  • A complete two-sided journey, working end to end — capture, clinical documentation, patient summary, sharing, prep, discharge — not a proof-of-concept for one step.
  • 1h 41m consultation documented in under 5 minutes. Real session, real length, real output.
  • Multilingual, code-switched transcription that holds up on actual South Asian consultations, which no existing consumer health app handles today.
  • Compliance treated as a feature, not paperwork — immutable audio archive built to DPDP Act 2023 from the start, because a clinical memory layer nobody is allowed to deploy is worth nothing.
  • A React Native app usable on real phones, not a scripted demo path.

What we learned

The documentation burden and the patient comprehension gap look like two separate markets. They are the same missing layer, viewed from two sides of the same desk — and the moment you capture the consultation once and serve both, the value stops being additive and starts compounding. Every session makes the next one better informed, for both people in the room.

We also learned that multilingual, code-switched support is not localization to bolt on in phase two. It's a first-class constraint from the transcription layer up. Get it wrong and the product doesn't degrade gracefully — it breaks completely, for exactly the patients and clinicians who need it most.

What's next for PatientDrive

  • Ship clinical task extraction and the medical record analyser to GA, closing the loop from spoken decision → tracked task → completed follow-up.
  • Expand language and dialect coverage, deepening beyond the 10+ Indian languages live today.
  • Clinician-to-clinician continuity — summaries shared securely across a patient's other doctors, not just their family, so referral and second-opinion visits inherit context.
  • FHIR / EMR integration, including ABDM-aligned profiles, so the memory layer writes into the systems clinics already run on.
  • Clinic partnerships across India, Nepal, and the USA to get PatientDrive to the point of care, where the consultation actually happens.

Built With

Gemini · Soniox · React Native · Node.js · DPDP Act 2023–compliant secure storage · FHIR (in progress)

Built With

Share this project:

Updates