AI-Based Mock Interview System Using Natural Language Processing and Real-Time Feedback

Abstract: We have created a mock interview tool that attempts to address one particular problem: most interview practice tools pose the same questions to all candidates, but this is not how actual interviews work. Our tool analyzes a candidate’s resume, identifies structured data within it via a DeBERTa model, and then generates questions based on this data via a locally installed Mistral instance via Ollama, with voice interaction via WebRTC and Whisper. After each response, we evaluate five criteria: whether it was relevant, technically accurate, insightful, well-expressed, and confidently stated. Our resume analysis achieved 91% precision and recall, and average response latency was close to 320 ms, but we are naturally a little nervous about how much we should read into these metrics, even from a limited test pool.

Keywords: Artificial Intelligence, Large Language Models, Named Entity Recognition, Natural Language Processing, Voice Activity Detection, WebRTC