CrackIt
Description
Real-time AI interview system with WebSocket relay pipeline enabling live speech-to-speech conversation with adaptive follow-up question generation and multi-model response evaluation.
Motivation
Technical interviews are anxiety-inducing partly because candidates rarely get realistic real-time practice. CrackIt simulates the actual conversational pressure of a live interview — with an AI that adapts follow-up questions based on your answers rather than following a fixed script.
Real-Time Speech Pipeline Architecture in CrackIt
TANAY_MATTA
CrackIt implements a real-time speech-to-speech interview simulation pipeline using WebSocket relay architecture and OpenAI's Realtime API, enabling sub-200ms audio round-trips with adaptive follow-up question generation driven by a multi-model evaluation chain.
Audio is streamed from the client browser to a FastAPI WebSocket handler, transcribed via Whisper, then routed through a multi-model chain: GPT-4o-mini generates contextual follow-up questions and scores the answer, while Hugging Face classifiers add sentiment and confidence signals. Synthesized TTS audio is streamed back, maintaining rolling conversation context across turns.