Research · An academic story · 2019 – ongoing
Does talking to AI make decisions better, or only make people surer?
That is my current question, about trust in high-stakes human-AI decision support. Its other half: when should AI hand a person to someone who can help?
It builds on two pieces of my own work. My first-author study of stigma and play in sex education in India proposes how play can lower the emotional barrier to difficult knowledge; the manuscript is in preparation. And on AI Studio at Innovaccer, I designed the human-in-the-loop workflows that set what the AI does on its own and what a user must confirm. The three directions below are written as questions, each with what exists so far and what does not.
Origin · 2019
My first questionnaire had to hide its aim
In 2019, in a student project at IIT Guwahati, I studied how to get more people into fitness activities. Laziness is taboo, so asking about it straight would have skewed the answers. I built a two-part questionnaire that hid its aim, and proposed three behaviour-change concepts from it. That is where the measurement problem started: ask people about something they are ashamed of, and the asking changes the answer.
The thesis · August 2020
My thesis found schools could not carry sex education
My graduation thesis started in August 2020: design interventions to bridge the gap between the sex education Indian students get and what they need. The plan was quantitative. Schools kept declining, and the experts around the study kept saying the same thing: teachers are not equipped to carry this. Formal education was not the way, and that conclusion redirected the next five years. The same years planted a second thread, a lifelong obsession with Mars and what isolation does to a mind. It resurfaces below as the ICE direction.
The thread · No grand plan
Building something is one thing. Surviving the real world is another.
There was no five-year master plan connecting any of this. What there was: an instinct that kept resurfacing, that building something and having it survive brutal real-world testing are two different beasts. It showed up in 2020 in the Digital Wellbeing Evaluation Tool, a framework for checking how features affect physical, psychological and social wellbeing, built from interviews with 11 product designers and design students. The framework has not been tested yet. It showed up again in the evals behind AI Studio.
The instrument · research graph
Screened by agents, decided at three human gates
The research graph behind this agenda is a screening instrument with me at the gates. About 480 papers were screened in 3 weeks. Agents triaged abstracts and wrote structured reviews against fixed criteria, and a critic shadowed every stage. Three human gates decided what entered the graph and what earned a full read. Standing infrastructure audits the result: integrity checks on every node, a weekly drift monitor, and weekly adversarial spot-audits. The full pipeline lives in the research graph; the reading log lists every source with citation and DOI.
Direction 1
Can rumination and mind-wandering be measured honestly, and reduced, in isolated, confined and extreme settings?
The target is default mode network activity, the self-referential loop behind mind-wandering and rumination. The hypothesis is regulation, not suppression: play that gives the loop somewhere structured to go, and a way to test whether it does.
The plan is to measure in the field, not the lab: EEG or fNIRS with experience sampling and validated mood scales, scans before and after a deployment, attention and mood as indexed by those instruments. fNIRS reads only the cortical surface, so the protocol treats it as a partial window.
The constraint is access: analog habitats, Antarctic stations, and the national programmes that run them decide who measures anything. The written protocol in the research graph is designed against that.
What is missing is named too: the field-methods training for EEG and fNIRS deployment. The current step is finding the lab and programme to host it.
Direction 2
Can play surface honest answers on topics people won’t talk about?
Surveys and interviews provoke the very stigma they try to measure. The question is whether gamified micro-learning can give people a private, non-judgmental space where honest signals surface as choices instead of confessions. What exists so far: the 2020 thesis no school would host, and the first-author paper that grew out of it, in preparation. The measurement study is future work, and it goes through ethics review before a single datapoint flows.
Direction 3
Can an AI tutor stay fair across cultures?
Long-horizon AI tutors need guardrails that hold pedagogical neutrality across socio-cultural contexts, not just across prompts. What I bring so far is design practice, not a study: the guardrails and evals I designed for AI Studio.
The paper · 2026
The writing has started
One manuscript, in preparation: my first-author study from the thesis. It is multi-method, with expert interviews, workshops, focus groups and surveys, synthesised into 77 insights. From them it proposes five design moves to lower the emotional barrier to a stigmatised subject.
Teaching · Writing
The students set the bar
Three courses at The Design Village as visiting faculty, May to November 2023. Visualisation and Representation was a course of my own; I co-facilitated Interface Design for Interactive Devices and Design Methods and Theory-I. The students went the extra mile on every detail, and the bar they set has stayed with me. Industry is where I collect what is worth teaching; the classroom is where it gets tested.
Learning and relearning. Always!
On the writing side, two essays. Agile Collaboration with Mid-Fidelity is already published on the Innovaccer Design blog, on how mid-fidelity keeps a team moving without over-committing too early. Stop Shipping “Finished” AI, the companion to the AI Studio story, is what a 1990s printer teaches about shipping AI products before they are done learning.
The base, in numbers
The graph is open
One thing the graph keeps insisting on: sexual-health self-report is unreliable. So the honest signal has to be behavioural: what people do in play, not what they say in surveys.