AI Meet
UX research teams are bottlenecked by researcher availability — every interview needs a human to moderate, schedule, and synthesize. As research scales, so does headcount. We asked: what happens if you remove the moderator from the room entirely?
The Need
A typical qual study involves 8–12 participants, each requiring a 45-minute block in a researcher's calendar. Then another block for synthesis. At scale, this is a fixed headcount problem — more research means more researchers.
The ask: make it possible to run 50 concurrent interviews with zero researcher time per session, while keeping the depth and nuance that makes qual research valuable in the first place.
The Key Insight
The AI is a participant in the call — not a post-processing feature.
The agent joins the WebRTC session via Stream's OpenAI Realtime bridge. It can hear, respond, pause when interrupted, and adjust based on what the participant says — in real time. From the participant's perspective, the experience is indistinguishable from talking to a human interviewer.
No account, no download. Participants click a link, enter their name, and join — a guest WebRTC token is issued on the spot.
When the call ends, a background job fetches the transcript, runs synthesis via agent-kit, and writes a structured markdown report to the DB. The researcher wakes up to findings.
How It Works
Sets agent instructions, voice, model, temperature. Generates a shareable per-booking link /b/{id}.
One booking per participant. System creates a Stream call and stores the meeting URL.
Enters name → system issues a guest token → participant joins the Stream WebRTC call. No account needed.
Stream webhook fires on session_started. AI agent joins via OpenAI Realtime API. Custom instructions, voice, temperature applied.
On call end: transcript JSONL fetched, Inngest summarizes with agent-kit (gpt-4o), markdown summary written to DB.
Reads AI summary, full transcript, and recording. Research workspace surfaces findings, quotes, and patterns.