What it is
An LLM interviewer for people with AI anxiety: knowledge workers who feel the ground moving and can’t see their way forward. You talk to it; it builds a 36-facet picture of how you actually work, then keeps the conversation going as a companion. Built with more discipline than anything else I shipped solo that year: ADRs, an eval harness for the scoring, per-user spend caps that degrade softly instead of blocking, moderation on every message.
Why I built it
The AI-anxious are real and mostly unserved: everyone sells them courses; nobody helps them understand themselves first. A questionnaire felt dead. An AI interviewer that over-focused on assessment felt clinical. I rebuilt it around the person. Typing felt slow. I rebuilt it around voice.
Why it’s paused
Voice created a real dilemma I haven’t solved. Voice models are fluent but transcribe poorly, and I made the mistake of showing their transcripts. The right design shows an animation while tools work in the background, but I still want a clean transcript underneath, because the deep analysis belongs to text models. Getting both (fluent conversation and a faithful transcript) without making everything slow is an open problem. It’s online and usable, in an honestly weird state, until I find the answer.
What I learned
Every interface rebuild was the same lesson: the product is the conversation, not the assessment. And discipline (evals, caps, ADRs) is cheap insurance on LLM products: the costs and failure modes you don’t bound will find you.