A Multimodal AI-Based Adaptive Mock Interview System

Technical interviews require candidates to solve problems, articulate their thinking and manage time constraints, but most current practice tools consider only one of these aspects. This paper reviews the development of AI interview systems from feature-based analysis of recorded interviews, to multimodal behavioural analytics and interviews driven by large language models (LLMs), to rubric-aware information elicitation. Related work includes talent analytics, resume matching, mock-interview platforms and speech, facial and text assessment. Comparison highlights a consistent gap. Screening, interactive questioning, multimodal feedback and evidence-based scoring are mostly studied separately. To address this, the paper proposes a web-based adaptive mock-interview framework. It builds a role-specific rubric from a resume or skill list, scores answers semantically, extracts speech features, optionally analyses facial-landmark trends, updates an evidence state for each rubric dimension after every response and selects follow-up questions where evidence is thin. It formally defines question utility, semantic scoring, evidence updating, speech rate, filler ratio, landmark normalisation, multimodal fusion and reliability metrics. It outlines an evaluation protocol for a planned within-subject study of adaptive versus fixed-question interviews with 30–50 computer-science students, along with fairness, privacy and ethics safeguards. The paper reports no experimental results. Figures cited from other studies are literature values only and the framework is a research design for interview practice, not a tool for hiring decisions.