The screening is a conversation. The 113-question form fills itself.

Most health research is limited by how fast answers can be written down, not by how fast they can be asked. This is an account of a study where the asking was quick and the writing down took far longer, and what we built so the conversation could do both.

The problem underneath the form

Dr. Pradip Dankhara's study screens school children in Gandhinagar district. Each child's record has 113 questions covering the school, the household, daily routine, height and weight, and a medical examination. The conversation with the child happens in Gujarati.

The old process had two halves. First, sit with each child, talk, and write notes on paper. Then, later, open the study's online form and retype all 113 answers, once per child. The screening was the part that mattered. The retyping was where the time went, and every answer copied from paper was another chance to copy it wrong.

So the brief was not really "build an app." It was to stop anyone having to write the same child down twice.

What we built

MedRecorder lives on the screener's phone. The screener talks to the child in Gujarati, exactly as before. The app listens, turns the speech into text, and fills in the questionnaire while the conversation is still happening. The screener checks the record, confirms it, and it goes into the study's existing Google Form, which files it into the same response sheet the study was already using.

Nothing downstream had to change. The study kept its form, its sheet and its way of analysing the data. The only thing that disappeared was the retyping.

A school visit, start to finish

One visit covers one school. The school-level questions are answered once, at the start, and copied into every child's record, so nobody answers them five hundred times.

Children arrive in whatever order the teacher sends them. The class is asked per child, with the visit's starting class already filled in, and once a visit covers more than one class the list can be filtered by class.

The microphone sits in a bar at the bottom of the screen, so it keeps recording while the screener scrolls, reads or fills something in by hand. A small counter beside it shows what the conversation has just filled, and tapping it jumps to any of those answers.

The medical examination starts from normal findings, so the screener only taps what is abnormal. Height and weight can be spoken or typed in whatever unit is at hand, centimetres, metres, feet and inches, pounds or grams, and the app stores them consistently and calculates BMI.

When a record is nearly done, the app names the next required question still unanswered and jumps straight to it. Confirming a child opens the next one waiting.

A whole class register in one photo

Names and measurements for a whole class are usually already written in the school's register. So the screener photographs the register, page by page, and the app reads every student row into the visit before the first child sits down.

Each page is read on its own so a crowded page gets full attention. The app also counts how many children it expected to find, and if it read fewer, it says so and names the page to photograph again. We added that check after an early version quietly used two of five selected photos and dropped the rest. A missing child is worse than a slow one.

Built for schools with weak signal

Schools in the district do not all have good mobile coverage, so the app was built to work without it. Every record is saved on the phone first. Typing answers, taking register photos and queueing finished records all work offline, and records send themselves once the phone is back in range.

Speech needs the internet to be turned into text, so recordings that fail to upload are kept on the phone and retried when the connection returns. An early version simply lost them on a weak signal, which meant the screener had to type what they had already said. That was the first thing we fixed.

Getting the district right

The study's form is the single source of truth. The app is built from an exact copy of its questions, their wording, and the answers each one allows, and it checks itself against that copy. If the form and the app ever disagree, if a required question would be sent empty, or if the app tries to send an answer the form does not offer, the checks fail before anything reaches the study. When the study updated its list of districts, schools and mid-day meal answers, the app followed.

Getting the questions right was not enough on its own. Early on, the AI was filing families who farm their own land as skilled workers instead of farmers. In a farming district, that would have tilted the study's data about household occupation. We caught it and corrected it before it reached the sheet.

Where the person stays in charge

The app fills what the conversation clearly supports and marks anything it is unsure of for the screener to confirm. Anything the screener types themselves is never overwritten. Every answer the AI filled keeps the words from the conversation that justify it, so the screener can see why an answer is there before accepting it. Nothing goes to the study until a person has checked the record.

Children's records stay on the phone until they are sent. Audio is used only to produce the transcript and is deleted as soon as the record is confirmed. The app sits behind a password, and it carries no analytics or tracking of any kind. A record already sent to the study cannot be deleted by accident.

What changed for the study

MedRecorder is now in regular use. More than 500 children across a dozen schools have been screened with it. The retyping step is gone: the answers are written down once, during the conversation, and checked once. Because no answer is copied a second time, the errors that copying used to introduce do not have the chance to appear. We estimate that the time spent on paperwork per child has fallen by around 70%.

The study also changes as it goes: new schools, new answer options, new requirements. Those changes reach the app quickly.

It is very easy to use, and it saves a lot of time and manual effort. When we need a change in the application, it is done in a couple of minutes.

If your team collects the same information twice, once on paper and once in a system, that second step is usually the one worth removing. The call is an hour and it is a working session, not a pitch. Bring the form you retype most often.

Schedule a call