Four exhibits, from the working files
The questions were audited before the users were.
Eight months of diabetes-care research in India. Before anyone was asked anything, every survey item, interview question and problem statement was scored for bias on seven axes, rewritten, and scored again.
Read this first
The analysis is deep. The sample is small.
Twelve structured interviews. Five moderated usability participants. Three card sorters. Against that: 8,176 rows of coding, mapping and scoring. Volume of analysis is not the same as weight of evidence, and conflating the two is the most common way research decks mislead.
So: the instrument design below is rigorous and reproducible — that part is strong. The behavioural findings are directional, not representative. Everything on this page is tagged for which it is.
Exhibit 01 · Reliability check
Aditya_-_Reliability_Check.xlsx · 4 sheetsEvery question was scored for bias before it was asked.
Seven axes, scored independently, each with a written reason. Then the item was rewritten and scored again. Pick a specimen — the scores and reasons below are the originals.
| Instrument | Items | Mean bias → after rewrite | Zeroed |
|---|
Where the bias actually was
Framing and social desirability, not stereotyping.
Cumulative weight per axis across all 251 items. The failure mode wasn’t offensive language — it was questions quietly telling people which answer was the good one. That shaped how the interview guide was rewritten.
105 interview questions went in biased. All 105 came out at zero.
The instrument is the reproducible part. What it found is not.Exhibit 02 · Secondary research & analysis
Aditya_-_Secondary_Research.xlsx · Aditya_-_Analysis.xlsxCollecting is easy. Cutting is the work.
Two consolidation passes ran in parallel — one on how people behave, one on what products already do. Both existed to get to a shortlist small enough to design against.
Behaviour → attributes
Qualitative influential attributes — behaviours, habits, aptitudes, attitudes — mapped per feature, per competitor, per age band, then deduplicated into a working set.
Market → features
Four competitor apps taken apart feature by feature, benchmarked against 60 quantitative criteria, then merged with features extracted from store reviews.
And then the shape of it
107 journeys, each assigned a flow type with a written reason.
Not a house style applied uniformly. Hub-and-spoke where users needed to return to a dashboard between actions; wizard only where a sequence genuinely couldn’t be reordered.
Exhibit 03 · Card sorting
Aditya_-_Analysis.xlsx · Card Sorting · 954 rowsUsers moved the menus. 63 changes went in. 22 didn’t.
Three participants walked the full menu tree, question by question. Every response was logged with a decision attached — actioned, or not. The ones that weren’t are the more interesting half.
Actioned
63Not actioned
22The rationale column was filled in for all 63 actioned items and none of the 22 declined ones. So the record shows what was turned down but not why — a real gap in the audit trail, and the reason the same sheet in later projects has a required-reason field on both outcomes.
Exhibit 04 · Moderated usability testing
Aditya_-_Usability_Testing.xlsx · 94 task observationsThree features tested. One cleared the bar.
Five participants aged 26 to 66+, timed sessions, 16 scripted tasks. Each observation carried a moderation question, an expected outcome, an acceptance call, a satisfaction rating and a recorded next step — seventeen fields per row.
What participants actually said, unedited
These are the improvement notes as written during the sessions. They are blunt, and that is the point — a wireframe that tests at 72% acceptance has told you something specific, and the notes are where the specifics live.
The corpus
Four workbooks. 37 sheets. 8,176 rows.
Everything above came out of these. One square below is twenty rows of coding, mapping or scoring — the whole corpus, at proportion.
Close
Research is done. The product is being built.
Four surfaces on one data spine — patient, caregiver, mentor, and a doctor/lab workspace — with the decisions locked per cohort before anything was drawn. Built to be correct when silent.
The instrument design is the reproducible part. The findings are directional, and labelled that way throughout.
Back to the case study