Authors: Assistant Professor B.Saritha, B.Sai keerthana
Abstract: Organisational development relies on recruiting and applicant evaluation. However, traditional interview techniques include manual coordination, static questionnaires, and subjective evaluations, which prolong the recruitment process and lead to inconsistent candidate selection. For screening candidates, arranging interviews, evaluating responses, and reporting performance, current recruiting systems often rely heavily on human engagement, which makes them inefficient when dealing with large amounts of applications. In addition, recruiters and applicants are both severely impacted by a lack of transparency caused by delayed response and restricted automation. In order to circumvent these restrictions, this research exposes an AI-driven online interview platform that automates the whole interview process using smart software agents and cutting-edge AI methods. With the suggested platform, a unified recruiting environment may be achieved via applicant registration, administrator permission, question development using AI, automatic response assessment, performance monitoring, and score alerts sent via email. The system uses a combination of Artificial Intelligence (AI), Bidirectional Long Short-Term Memory (BiLSTM) models, Convolutional Neural Networks (CNN), Speech-to-Text (STT), Text-to-Speech (TTS), and Natural Language Processing (NLP) to analyse candidate responses during dynamic interviews and assess sentimental and emotional traits. Intelligent interview management and objective candidate assessment are accomplished through the collaborative efforts of a multi-agent architecture that includes the Question Management Agent (QMA), Response Management Agent (RMA), Multimodal Emotion and Sentiment Analyser Agent (MESAA), and Comprehensive Evaluation and Scoring Agent (CESA). Emotion identification, sentiment analysis, and question management are all handled by the framework using datasets such as SAVEE, TESS, CREMA-D, ISEAR, and MedMCQA. The metrics BLEU, ROUGE, BERTScore, and Completeness Score are used to assess the candidates' answers. In contrast to the CNN model's 95% training accuracy and 67% validation accuracy for audio emotion identification, the BiLSTM model only manages 70% training accuracy and 60% validation accuracy when it comes to text emotion recognition, according to experimental assessment. The suggested platform offers a scalable, efficient, and impartial answer to contemporary recruiting by combining intelligent automation with thorough applicant assessment; all while maintaining the original methodology, architecture, datasets, and trial results.
