AI Interview Assessment System

Abstract

Automated evaluation of human responses has become increasingly important in intelligent recruitment systems and digital skill assessment platforms. Advances in natural language processing enable semantic understanding of textual responses beyond keyword matching. This study presents a web-based interview assessment system that integrates semantic similarity modeling using transformer-based sentence embeddings to evaluate candidate responses against predefined reference answers. System is built using Flask framework with SQLite database for structured storage of users, questions, sessions, and response analytics. A pretrained allMiniLM-L6-v2 sentence-transformer model is utilized to encode both candidate and model answers into highdimensional vector representations. Cosine similarity is computed between embeddings, and a hybrid scoring strategy combining maximum similarity and average similarity is applied to generate normalized scores on a 0– 10 scale. Administrative modules provide topic management, question curation, and performance analytics, while candidate modules support real-time interviews with progress tracking and feedback generation. Browser-based speech-to-text capability enhances accessibility. Experimental evaluation demonstrates that semantic-based assessment improves robustness over traditional keyword-based methods, enabling consistent and scalable evaluation of technical interviews. Future enhancements may incorporate multimodal response analysis, domainadaptive fine-tuning of transformer models, and advanced explainable AI techniques to further improve fairness, interpretability, and assessment accuracy in automated interview systems frameworks and services.