PrepWise AI

The problem

Text-based interview practice never rehearses the thing that actually breaks candidates: speaking answers out loud, under time pressure, to something that responds.

What I built

A Next.js application that runs live voice interviews through Hume's real-time voice API. It parses the user's résumé against a target job description to generate hyper-personalized questions, then scores answers into structured progress feedback. Clerk handles authentication, Drizzle and PostgreSQL persist sessions and analytics, the Vercel AI SDK drives Gemini for question generation and grading, and Arcjet shields the endpoints from bot abuse.

Impact

Tracks complete practice history, logs session metrics and converts spoken answers into structured, scored progress feedback.

Architecture

Next.js 15 App Router with React 19 and Tailwind v4. Drizzle ORM over PostgreSQL for sessions, questions and analytics. Hume's voice React SDK handles the real-time audio loop; the Vercel AI SDK abstracts Gemini for generation and scoring. Arcjet provides bot detection and rate limiting at the edge of the API routes.

Key features

• Real-time voice interviews with an adaptive AI interviewer
• Résumé parsing matched against target job descriptions
• ATS compatibility scoring and keyword optimization
• Session analytics with scored answers and progress history
• Bot and abuse protection via Arcjet