Technical Report on KshemaGPT: A Multi-Agent LLM for Agriculture & Enterprise AI

Rural India is highly diverse in terms of customs, languages, and literacy. With more than 22 officially recognized languages, communication and information exchange remain a key factor. Although, the rise of open-source Large Language Models (LLMs) has significantly filled the gap in generic information availability, the need for domain-specific information remains unquenched. Particularly policy and stakeholder specific information in the insurance domain presents a critical challenge owing to its less awareness among the rural population. In this work, we demonstrate a multi-agent architecture which brings in multiple domain-specific agents such as Policy Agent, Crop Agent, User Agent, Employee Agent, Translation Agent, and Speech Recognition Agent combined through Query Router Agent to handle the diverse input and information requirements. We have deployed the multi-agent architecture in Azure Cloud while considering various aspects such as security, scalability, and observability. We have conducted a latency analysis to evaluate the execution times of each agent for one and five simultaneous users. Further, we have also evaluated the efficiency of Query Router Agent in terms of classification metrics.