Voice Interview Coach

■Problem

Practicing for interviews alone doesn't feel like an interview. Reading questions off a list skips the hard part: answering out loud, in real time, while someone asks follow-ups.

Voice Interview Coach runs a spoken mock interview in the browser. You pick a role, an interview type (behavioral, technical, or mixed), and a difficulty, then talk with an AI interviewer that asks one question at a time and follows up on your answers. It doesn't coach you mid-interview. Evaluation happens afterward, when a separate model reviews the transcript and returns overall feedback, strengths, areas to improve, communication notes, and feedback on each answer.

■Role

I designed and built it on my own: product scope, architecture, the realtime client, the feedback engine, and the test suite.

■Stack

• Next.js + React + TypeScript for the app and its two server routes.
• OpenAI Realtime API over WebRTC for the live conversation. Audio streams directly between the browser and OpenAI, which keeps latency low enough to feel like a real conversation.
• A server-issued, short-lived credential for each session. The permanent API key never reaches the browser.
• A separate feedback model with a strict JSON schema for post-interview evaluation. Keeping the interviewer and the evaluator separate means the interviewer can stay in character.
• Vitest for 132 unit and integration tests.

■Challenges

Keeping the transcript correct. Realtime events don't arrive in a tidy order. Two turns can finish transcribing out of order, a candidate can interrupt the interviewer mid-sentence, and the text for an interrupted turn can arrive before or after the interruption signal. I wrote the transcript as a pure reducer over raw Realtime events. It orders turns by the API's previous_item_id rather than arrival time, and it never lets a late event overwrite a final status. That made each edge case a fast, deterministic test instead of something only reproducible in a live browser session.

Not losing the last answer. Ending an interview while the candidate is still speaking was the most persistent source of bugs. The final answer could vanish because the server hadn't yet told the client about it. The fix was a short drain window after a live disconnect: if anything is known to be pending, the app waits for it. Anything that still can't be captured is marked unavailable rather than silently dropped, and both the evaluator and the results screen disclose that feedback covers an incomplete interview.

Not trusting the model's output. Even with a strict schema, feedback is validated like any other untrusted input. Each piece of per-answer feedback must reference a candidate turn the model was actually shown, with no duplicates, and scores must stay in bounds. Otherwise the whole response is rejected, so a hallucinated turn can never be presented as feedback on something you said.

■Status

It runs locally with your own OpenAI API key; there's no hosted demo yet. Accounts and saved interview history are deliberately out of scope until the core interview experience is solid.