AI audience that grills you after you present your work

A real-time voice flow where a person stands up and presents their own work out loud, the way they would to a manager or a client. An AI audience listens to the whole thing, then runs a live Q&A, asking follow-up questions tied to specific points the person actually made (not generic questions), pushing back where their reasoning was thin. The session closes with a structured feedback report naming what the person did well, what landed, what didn't, and what to push on next time, delivered conversationally instead of as a wall of text.

Most ways of checking whether someone understood their own work rely on what they wrote or clicked. Presenting it out loud and then defending it under real follow-up questions is a much higher bar: you have to explain it in your own words, then hold up when someone pushes on the weak point. This project builds toward that moment, so it has to handle live questioning without breaking the illusion of a real audience. If the AI fumbles on the third question, this stops being useful as real practice.

A LiveKit voice session split into four phases. Intro: the AI greets the person, sets the topic, and signals it's listening. Presentation: the person delivers their piece, and every utterance is saved per turn with a timestamp and phase tag, so a mid-session crash never loses what was said. Q&A: once the person wraps up, the AI reads back the full transcript, picks the points that need pressure (claims that were thin, jumps that weren't justified, choices made without showing why), and asks targeted follow-up questions in voice. The person can defend, hedge, or admit they don't know, and each answer shapes the next question. Feedback: the AI generates a structured report covering strengths, weaknesses, and stretch areas, then delivers it conversationally rather than dumping a list. Deepgram handles speech-to-text, ElevenLabs handles voice with streaming text-to-speech so the AI starts replying before it's finished composing the full sentence, and the whole session runs as a LiveKit room over WebRTC.

Sole author. Built the four-phase machine (intro, presentation, Q&A, feedback), the per-utterance transcript persistence, the live Q&A generator that pulls questions from what the person said, and the end-of-session feedback report.