Voice-first first-time onboarding

A voice-first onboarding agent that runs in the first few minutes a user lands on the platform. It seeds the user's profile (goals, prior experience, and how they frame problems) from how they talk about their own work, not from a typed survey.

The first five minutes on the platform are the highest-leverage moment to learn who the user is. A typed survey gives what they think they should say. A voice conversation gives how they actually talk. The platform cares about the second one, so onboarding is built around it.

A LiveKit voice runner with Gemini handling the conversation directly. As the user talks, every utterance is shipped in the background to the calibration ingest endpoint without waiting for the reply to finish, so the platform is already learning from them before they have finished introducing themselves. Once the conversation closes, an initial-calibration module reads those early entries and pulls out what was said about goals, prior experience, and how the user frames problems, then writes a seed profile into the platform's running user model. The first real session the user lands in already has shape behind it instead of starting from a blank slate.

Sole author of the onboarding flow end to end.