Building a user research tool fast enough for product teams, and rigorous enough for researchers.
Enterprise-grade user research costs big money, the kind smaller product teams don't have. UX4.ai was HFI's bet that AI could close that gap: a platform that combines real user interviews with synthetic AI participants, backed by 40 years of HFI's research methodology. I worked directly with leadership on this internal venture.
I came in at zero and stayed through our first customers; Samsung and TVS Motor. In terms of product strategy, I benchmarked 12 platforms to map the market, shaping how we positioned UX4's hybrid model against competitors.
My cold outreach surfaced an insight: research-inexperienced orgs weren't buying software, they were buying insights. That made our software + service engagement the sharper pitch.
I worked on features and UI, this is when I started shipping code and understanding repos. Highlights follow:
Designing filters around how researchers actually synthesize
Researchers stated they don't read findings linearly. They drill down, cross-cutting between personas, segments and themes until they've narrowed to the participants worth reading. The existing UI gave them everything at once with no way to narrow. I designed a filtering system that lets users layer various filters progressively, arriving at precisely the participant slice they want to investigate.
Making chat more accurate
Our chat feature was unreliable at multifaceted questions. Ask it about specific themes, segments and their interrelation and it would frequently pull from a broader slice of data, or generalise across the dataset when you wanted a narrow answer. To counter this, I designed an inline tagging system that lets users reference personas, segments, or themes in conversation, so the AI knows exactly what transcripts to look at.
Facilitating synthetic focus groups
Focus groups are dynamic, multi-person conversations, but our first interface was a plain chat thread, every reply stacked one after another, harder to track with each new one. I redesigned it to keep the stimulus pinned in view alongside a scannable grid of responses (each with a sentiment indicator), so users compare reactions at a glance instead of piecing them together from a scroll.
Testing the efficacy of synthetic participants
To validate UX4's core claim, I ran a comparative study using the same protocol on both synthetic and real participants. I analysed the findings and wrote an article on how synthetic participants can compliment human data for user research.