TruInterview: Cheat-Proof Voice AI Screening

Technical hiring today means choosing between two bad options: expensive proctoring vendors billed per candidate, or generic MCQ tools that never test how someone actually thinks through code. TruInterview replaces both with one self-hosted platform. Recruiters drop a role description into an invite link; TruInterview classifies it, generates role-specific MCQ and open-ended questions, and runs a live spoken interview so the candidate never touches a form. The voice layer is fully self-built: browser TTS plus a local faster-whisper STT microservice with voice-activity detection, so candidate audio is transcribed on our own server, never sent to a third-party call vendor. Meanwhile, browser-native proctoring locks down fullscreen, multi-monitor, clipboard, and DevTools access, and AI vision checks webcam frames and screen captures for face and focus anomalies, so cheating gets caught without a human proctor. Every AI task runs on Fireworks AI, with each task routed to the model built for that job: DeepSeek V4 Flash drives live interview chat where every second of latency costs candidate trust, MiniMax M3 reads screen-share frames mid-conversation, Qwen3-7B Plus handles proctoring vision and resume PDF extraction, DeepSeek V4 Pro generates questions and matches role taxonomy, Kimi K2-7B Code rewrites resume content, and GLM-5.2 compiles the final evaluation scorecard. One config file maps every task to its model, so tuning speed against reasoning depth per task is a one-line change, not a rewrite. TruInterview scores communication, problem-solving, and code quality automatically, so recruiters get a decision-ready report in minutes, and candidates get an AI resume optimizer tuned to the same role. Stack: PHP, PostgreSQL, vanilla JS, Fireworks AI - no build step, no framework lock-in, deployable anywhere.