NexterView
NexterView is an AI mock interview platform. Generate a customised interview for any job role, answer with your voice, and receive a structured performance breakdown with skill-level scoring.
The Voice Problem
The Web Speech API has three failure modes that compound each other: transcription drops mid-sentence, silence thresholds misfire and end the answer too early, and partial results arrive out of order. A voice interview session that hiccups or cuts the user off mid-answer isn't just technically broken, it's anxiety-inducing, which is exactly the wrong experience for an interview practice tool. The fix required buffering partial transcripts and only committing an answer when a confidence threshold was met, with a grace period before the session interprets silence as "done." Recovery from dropped recognition sessions mid-interview was handled by reattaching the SpeechRecognition event listeners without restarting the question, so the user sees nothing when it recovers.
Built using Next.js, TypeScript, Tailwind CSS, Prisma, Google Gemini AI, Clerk, Arcjet, PostgreSQL, Web Speech API
Key Decisions & Outcomes
- Voice-to-text interview sessions built on Web Speech API with partial transcript buffering and confidence-threshold answer detection to handle mid-sentence drops and silence misfires without interrupting the user
- Question generation via Gemini 2.5 Flash is customised by job role, seniority level, and focus area, so every session produces a unique set of questions
- Scoring produces a skill-breakdown report (communication, technical accuracy, structure, confidence) rather than a single number to provide more actionable feedback
- Rate-limited with Arcjet to prevent abuse of the AI pipeline in production without adding friction for real users
If you're curious, feel free to explore: