Introducing the M3D-MoA Framework and Document Augmentation to Enhance In-Context Learning for LLMs: ERIC, a Waste Classification Case Study.

In this thesis, I present the development and implementation of ERIC (Environmental Resource Intelligent Companion), an innovative AI-driven solution I designed to enhance environmental management practices within Amazon's Environmental Assurance and Protection (EAP) department. My thesis focuses on leveraging Large Language Models (LLMs) and advanced prompt engineering techniques to address complex challenges in waste classification and environmental compliance. I introduce two novel frameworks: the Multi-Model, Multi-Document, Mixture of Agents (M3D-MoA) framework for orchestrating multiple LLMs, and the Document Augmentation technique for enhancing In-Context Learning. These innovations address key limitations in current LLM applications, such as context window constraints and resource intensiveness. I outline a systematic Gen-AI Application Prototyping process, detailing the development of ERIC from concept to pilot implementation. My thesis includes a comparative analysis of LLM enhancement techniques, justifying the selection of In-Context Learning for the project. The UK pilot implementation of ERIC demonstrates significant improvements in waste classification efficiency, with a 75% reduction in processing time and 93% accuracy in EWC code assignments. I present a preliminary Return on Investment (ROI) analysis indicating substantial potential for cost savings and operational efficiencies. I conclude by discussing the implications of AI integration in environmental management, highlighting areas for future research, and emphasizing the transformative potential of AI-driven solutions in corporate environmental compliance. My internship project and innovations contribute to the growing body of knowledge on practical AI applications in environmental management and offers insights into the challenges and opportunities of implementing AI solutions in complex regulatory environments.