Overview

Doctors were spending more of each appointment typing than talking. Manorama EMR lets them talk instead: the doctor speaks through the consultation, an in-house medical AI model transcribes it, and the record fills itself in while the doctor stays with the patient. I designed it end to end over five months, from the first meetings with senior doctors, through four rounds of wireframes they rewrote, to the desktop and mobile screens now used in clinics and hospitals.

Details

Role: Lead product designer, Research to final UI
Timeline: 5 months, shipped April 2025
Team: Manorama tech team and a panel of senior doctors

Understanding the Problem

Fifteen minutes of typing for every patient.

In the old system a doctor typed every complaint, finding and prescription into dense forms while the patient waited across the desk. Documentation alone took 15 to 20 minutes an appointment. That capped how many patients a doctor could see, and it pulled their eyes from the patient to the screen. We started with meetings between the tech team and senior doctors, and the problem turned out to be four problems at once.

Walking through the existing EMR with the tech team and doctors:

Research, Strategy and Competitive Analysis

Voice AI already existed. Our job was to make doctors trust it.

I used Suki AI, the voice assistant clinicians use in the US, as our benchmark: how it captures a consultation, how it shows what it heard, and how a doctor fixes a mistake. It proved that voice documentation works. Setting it against what our doctors told us showed four places where we could go further.

Where we went further

Error correction: fix a wrong word with one tap or one voice command, without dictating again.

Built for low technical comfort: one screen in plain clinical words, usable with no training.

Voice and manual as equals: any field can be edited by hand as fast as by voice.

Our own medical model: voice built into the EMR itself, on an in-house model trained on medical language.

The key insight was that successful implementation required designing with doctors, not just for them.

Key Insight

Design with the doctors, not for them. Senior doctors reviewed every iteration from here on.

Process

Four iterations, each one rewritten by doctors.

V1 Foundation: the core voice interaction, the information architecture and the first documentation flow.
V2 Refinement: simpler navigation and a hierarchy that follows clinical priorities instead of technical logic.
V3 Optimisation: faster error correction, layouts built for scanning, and the first mobile screens. Testing showed which parts doctors really used.
V4 Final: hands-on testing with senior doctors and the final AI touchpoints. It looked nothing like V1, and that was the point.

Iteration 1: Foundation

Created initial wireframes with prototypes establishing the core voice-control interaction patterns, information architecture, and basic user flows for patient documentation.

Iteration 2: Refinement

After the first review with stakeholders, I simplified navigation based on usability feedback, refined the voice command structure, and adjusted information hierarchy to match clinical priorities rather than technical logic.

Iteration 3: Optimization

This version focused on improving error correction workflows, optimizing layouts for quick scanning, and enhancing visual hierarchy, also added mobile responsive designs. Testing revealed which elements doctors actually used versus what I thought they'd need.

Some more feedback sent by the Doctors' team as a PPT:

Key decisions

Four decisions that made voice work in a clinic.

Decision 1

Voice is the input; the screen is the proof.
The doctor talks and a narration panel shows the live transcript, while the record fills itself in beside it. Seeing each phrase land in the right field is what lets a doctor trust the AI without reading the transcript twice.

The doctor speaks once; vitals, complaints, history and diagnosis are filled in and ready to check.

Decision 2

Everything the AI writes is one tap from editable.
The doctors asked for full manual control, so voice and touch are equals. Every section is a tile with its own Add button, sections the AI filled are highlighted, and findings are color coded, positive in red and negative in green, so a doctor scans a record instead of reading it.

Decision 3

Laid out in the order doctors think.
Vitals run across the top, then complaints, examination and diagnosis, then orders and medicines. Services used in 20% of visits moved into Extras, so the 80% a doctor needs stay on one screen.

Decision 4

The same record on a phone.
Vitals, summary, prescription and completion as four tabs, with the voice button always in reach of a thumb.

Final designs

The shipped product.

Impact

From 15 to 20 minutes to 3 to 5.

75% less documentation time: from 15 to 20 minutes an appointment to 3 to 5, so doctors see more patients.
95% transcription accuracy, up from 90% at launch as the model learned, and the corrections left take seconds.
In use across several clinics and hospitals, with the strongest response from the senior doctors who had been most sceptical.
Doctors report more eye contact and attention in the consultation, because the record keeps itself.
Method: I sat in on consultations with the doctors and timed the documentation for each patient, from opening the record to saving it, first on the old EMR and again on the new one. Accuracy is the share of AI-transcribed fields the doctors kept without correcting.

Key Features

Voice-Controlled Documentation

Natural language processing captures doctor-patient conversations with medical terminology recognition. The system understands context, not just words, automatically populating relevant fields across EMR forms.

Intelligent Automation

The AI provides contextual suggestions based on symptoms and patient history, generates clinical notes automatically, and uses a continuous learning model that improved accuracy from 90% to 95% over time.

Intuitive Error Correction

Visual highlighting of AI-transcribed content with single-tap editing of any field. Large, accessible interaction targets and voice-based correction commands provide multiple pathways to fix errors quickly.

Streamlined Interface

Clean layouts prioritize essential information with logical grouping based on clinical workflows. The design accommodates low-tech-comfort users through progressive disclosure and clear visual hierarchy.

Reflection

The best AI feature is the one doctors stop noticing.

Three things I took from this project.

Iteration tells the truth: four rounds turned a well researched first wireframe into something quite different, and every change came from watching doctors use it.

The users are co-designers: senior doctors knew things about clinical work that no amount of observation would have shown me.

Simplicity is a discipline: for doctors with very different comfort with technology, every element had to earn its place. We set out to cut documentation time, not to show off the AI, and that is why the AI works.