This project applies advanced AI methods to one of the largest digitized federal environmental archives in existence: 155,000+ documents (620+ GB of PDFs) produced by the U.S. Environmental Protection Agency’s Superfund hazardous waste cleanup program, spanning 1977 to 2025. The project’s long-term goal is a book-length history of the Superfund program tracing its evolution, priorities, and human impact from 1980 to the present. Prior project phases established a document preservation pipeline, a benchmarked retrieval-augmented generation (RAG) extraction system using FAISS vector indexing and the Gemma 3-27b language model, and structured geospatial outputs. The Fall 2026 co-op advances the project into its most technically ambitious phase: building agentic AI systems capable of cross-site synthesis and historically meaningful inference.
Research objectives:
• Design and implement a two-tier agentic AI architecture in which an orchestrator LLM decomposes complex historical and regulatory questions into targeted sub-queries, executes them against the existing extraction pipeline, and synthesizes verified, cited analytical narratives across sites.
• Evaluate and benchmark orchestrator model candidates (distinct from the extraction-stage LLM), and deploy the full pipeline on the WIT Unity GPU cluster.
• Build inverse-query workflows capable of answering questions the corpus was never structured to answer directly; such as which industries generated the most contamination sites, how remediation strategies evolved by decade, and what patterns exist in cleanup timelines across communities of different demographic profiles.
• Implement confidence propagation and contradiction detection so that synthesized narratives carry trust indicators and surface documentary disagreements for historian review.
