AI-moderated interviews
An AI interviewer I co-founded at Torchbox, built to reach the people traditional research can't.
I co-founded an AI interviewer that's already reached people our usual methods never could.
Most user research still runs on scheduled calls and free time nobody has. Night shift workers, people on their feet all day, people who never get asked. I wanted a way to reach them that didn't lose what makes qualitative research worth doing in the first place.
Why I built it
Torchbox works with public sector and charity clients who need to hear from thousands of people, not dozens, and traditional interviews just can't scale to that. Surveys can, but they lose the why behind the numbers.
I couldn't find a tool that solved that properly, so I co-founded Sonar with Madhav Manoj: a voice AI interviewer that runs real research conversations at scale. Someone can talk to it from their phone, in a break room, at midnight after a shift, no scheduling, no Zoom link required.
Where it fits
Sonar doesn't replace human-moderated interviews, and it was never meant to. A skilled researcher in a real conversation still gets you things AI can't: the follow-up that comes from genuinely listening, the read of a room, the trust built over an hour together. There's a lot of value in interviews run by a person, and that isn't going anywhere.
What Sonar is for is the people that kind of conversation can't reach: shift workers with no time to speak to you, or the hundreds of stakeholders whose views need gathering when sitting down with each one simply isn't possible. It's a tool for those specific situations, not a stand-in for good interviewing.
How it works
Sonar is a large language model with a voice: it listens through speech to text, replies through text to speech, and runs the interview from a set of prompts I helped design. Those prompts cover tone, pacing and what to do if someone raises something serious.
We built it ourselves rather than buying an off the shelf tool, mainly for compliance. Most AI providers aren't set up for UK GDPR, and we needed full control over consent, storage and what happens to people's data.
Madhav Manoj, my co-founder, led the engineering. I've written more about how we built it and the early experiment that started it.
Where it's been used
At Guy's and St Thomas' NHS Foundation Trust, I ran Sonar alongside 61 traditional interviews, and it reached staff who'd never have made it into a scheduled research session. 85 more people took part this way: porters, matrons, night shift workers, community health staff.
The findings lined up closely with the human led interviews, and added things we'd have missed entirely. Staff with dyslexia asking for different content formats. Community workers using national guidance because the local version was impossible to find remotely.
The fuller story, including what changed as a result, is in my GSTT case study (password protected).
Getting it in front of people
I presented Sonar at the Local Government Technology Conference alongside GSTT's product manager, and it's since been shared at a CEO charity leaders' breakfast. It's also been written up as a model for AI assisted research in the public sector.
What's next
There's more to build. An admin interface would let researchers run and adapt interviews without touching code, and better synthesis tools would mean the output is ready to analyse, not just a pile of transcripts. I'd also like support for more languages, so people can speak in whichever one they're most comfortable in.
I'm also looking beyond research. Sonar could work for incident reporting, ongoing feedback or supporting a human agent on a live call.