DrukSerenity: A Culturally-Aware Multi-Agent Chatbot Framework for Mental Health Support in Bhutan

Abstract

Globally, mental health challenges represent a significant public health issue, and Bhutan similarly faces barriers to accessing mental health support due to cultural stigma and limited resources. Currently, there are mental health chatbots, but they are primarily for Western, English-speaking users. There is a need for a system that safely and culturally adapts for Bhutanese people. Therefore, this paper introduces DrukSerenity, a multi-agent chatbot that addresses crisis, offers birdirection Dzongkha-English translation and provides appropriate cultural content using data consisting of 977 Dzongkha proverbs ,alongside 22 cultural, policy, and clinical documents from related institutional sources. The system was evaluated on a synthetic benchmark of 542 messages using rule-based checks and an LLM-judge protocol, showing strong safety performance (100% crisis precision, 92.2% crisis F1) and high cultural fidelity (0.94 RAG fidelity), with an overall functional accuracy of 83.8%, where Emotion classification accuracy and processing latency remains a limitation. Although formal clinical validation with licensed practitioners has not yet been conducted, an informal usability pilot with 20 Bhutanese users received good ratings (overall experience 3.80/5) alongside concerns about proverb overuse. DrukSerenity nonetheless offers a replicable template for building culturally-aware mental health AI in other lowresource, linguistically underserved settings.