An Integrated AI Mock Interview System

Abstract

In the rapidly evolving digital recruitment landscape, the traditional methodologies of candidate assessment are being challenged by the need for scalable, objective, and high-integrity interview preparation systems. This study presents an integrated AI-driven mock interview framework that harmonizes large language models (LLMs) with real-time computer vision (CV) to simulate rigorous professional environments. The proposed architecture utilizes a PHP-based backend coupled with a MySQL relational database for structured data persistence of candidate profiles and performance metrics. To ensure session integrity without compromising user privacy, the framework incorporates a sophisticated multi-modal proctoring mechanism. This module employs MediaPipe Face Mesh for client-side, edge-based inference of 468 3D facial landmarks, enabling real-time gaze tracking and head-pose estimation without requiring server-side video storage. For dynamic conversational interaction, the system leverages the Llama 3.3 model, optimized via the Groq LPU (language processing unit) inference infrastructure. This integration maintains sub-500 ms response latencies, thereby preserving the natural cadence of human dialogue and mitigating the “uncanny valley” effect often found in high-latency AI agents. Furthermore, the system implements a dual-layered evaluation engine: a semantic NLP parser for technical accuracy and a multi-modal sentiment analyzer that correlates vocal pitch, textual intent, and behavioral cues. Experimental results, derived from a pilot study of 50 diverse professional resumes and 100 simulated sessions, indicate a 94% accuracy rate in distraction detection and a strong statistical correlation ($r > 0.85$) between AI-generated scoring and human expert assessments. The framework provides a privacy-conscious, low-latency, and scalable solution for mitigating interview anxiety while establishing a standardized benchmark for modern performance evaluation.