Enabling Large Language Models and Agentic Workflows on Legacy GPU Architectures for Building Scalable Farmer Advisory Systems in India

Large Language Models (LLMs) and agentic workflows are increasingly being applied and evaluated across multiple domains, including agriculture, where they support intelligent, context-aware advisory services. Such systems can assist farmers across the agricultural lifecycle, including soil preparation, sowing decisions, crop health monitoring, pest identification, irrigation planning, harvesting, and post-harvest management. In the Indian context, where agricultural practices vary significantly across regions and are deeply rooted in local knowledge, Generative AI and Natural Language Processing can be used to deliver timely, localized, and language-accessible agricultural advisories. Adapting LLMs with region-specific domain knowledge enhances their relevance and practical utility, particularly when delivered through chatbot-based interfaces that support local languages. However, training and deploying such systems typically require substantial computational resources, often relying on modern GPU architectures. This paper examines the practical considerations involved in leveraging existing computational infrastructure for Generative AI workloads and proposes an approach for their effective use in resource-constrained settings. The paper further presents a proposed system architecture incorporating agentic AI workflows for agricultural applications, illustrating how task-oriented agents powered by LLMs can support agricultural chatbot functionality. It also discusses key design considerations in integrating model outputs with agent-based workflows and constructing scalable pipeline architectures tailored to agricultural use cases.