A Privacy-Preserving AI Companion for Student Mental Wellness: Adaptive Conversational Support with Sentiment-Aware Monitoring and Retrieval-Augmented Guidance

Abstract: Psychological distress among university students has risen markedly, yet institutional counselling capacity remains limited and help-seeking is often deterred by stigma and waiting times. Conventional digital wellness applications typically rely on remote cloud services, which raises legitimate concerns about the confidentiality of highly sensitive emotional data. This paper presents a local-first, assistive software framework that delivers empathetic conversational support, continuous sentiment-aware monitoring, and document-grounded informational guidance entirely on institutionally controlled infrastructure. The platform combines a reactive single-page interface with an asynchronous service core in which all generative responses are produced by a locally hosted large language model, eliminating any transmission of student dialogue to third parties. Incoming messages are screened by a lightweight natural-language pipeline that fuses lexical polarity estimation with a curated distress-and-stress lexicon, allowing the conversational agent to adapt its tone and to escalate supportive prompts when crisis language is detected. Self-reported mood and stress check-ins have persisted and transformed into a composite well-being indicator, while an anonymised analytics module equips counsellors with cohort-level trends derived through linear regression. Curated wellness literature is made query able through a retrieval-augmented generation pipeline using sentence-transformer embeddings and a vector index. A real-time channel notifies designated staff when stress thresholds are exceeded. Evaluation shows millisecond-scale sentiment screening, sub-second time-to-first token for streamed replies, and a distress-recall of 0.93 on annotated samples. The framework offers a confidential, infrastructure-light foundation for scalable campus mental- health support and is positioned strictly as an aid rather than a clinical substitute.

Keywords: Student mental health, conversational AI, sentiment analysis, retrieval-augmented generation, large language models, privacy-preserving systems, emotion monitoring, well-being analytics