Pocket Interview Coach

People fail interviews less on skill than on nerves — rambling, saying "um", going monotone, or giving no concrete results — and text-based AI prep ignores how you speak, which is half of what an interviewer judges. Pocket Interview Coach is a voice-first mock-interview simulator: pick a role (plus optional company, job description, resume, persona, and difficulty) and it runs a realistic spoken interview. For every spoken answer it scores both dimensions — content (structure via STAR, specificity, relevance) and delivery (pace, tone, filler words, pauses, and volume steadiness, measured from your voice) — coaches you in plain English ("you sped up to 180 wpm — reads as nervous"), scores each answer 0-100, asks devil's-advocate follow-ups, and ends with a readiness report, a cheat sheet, and PDF/JSON exports. It runs two models for two jobs, like a real hiring loop: Gemma 4 is the coach that writes questions, scores content and delivery, asks follow-ups, and builds the cheat sheet, while a Fireworks model on AMD Instinct GPUs is the hiring manager that independently reviews the interview and makes a real Strong-Hire to No-Hire call with confidence and the case for and against you, and also rewrites your answer into a stronger version — two independent models, so the verdict is a genuine second opinion, not one model grading itself. This highest-stakes hire/no-hire decision runs on AMD Instinct GPUs via Fireworks, while local speech-to-text and prosody run on CPU with a switch to ROCm/AMD-GPU. Interview prep is a huge market, yet existing tools grade text, not voice and delivery — this is the 2am practice partner that grades not just what you say but how you say it, and whether you'd get the job. Built with Streamlit (containerized with Docker), Gemma 4 via an OpenAI-compatible API, Fireworks gpt-oss-120b on AMD, faster-whisper, librosa, Plotly, and fpdf2; deployed on Streamlit Community Cloud.s