LLM-Assisted Farm Management Insights
Built a human-in-the-loop LLM pipeline that converts agricultural research into actionable farm management recommendations, published in Scientific Reports.
Overview
Developed a system that uses multiple large language models with expert oversight to systematically screen academic literature and generate evidence-based soybean farm management plans.
Problem
Increasing food production sustainably requires translating a growing body of agricultural research into practical guidance for farmers—a process that is slow, labor-intensive, and difficult to scale manually.
Approach
Designed a multi-stage pipeline: systematic literature search using the PICO framework on Web of Science, parallel screening by four LLMs with expert arbitration, extraction of evidence and relevance assessments by the two top-performing models, inconsistency detection, and final LLM-generated management plan synthesis.
Constraints
- Recommendations must be grounded in peer-reviewed research, not LLM hallucinations
- Multiple LLMs needed cross-validation to ensure screening accuracy
- Expert arbitration required at every stage to maintain scientific rigor
Key Decisions
Multi-LLM consensus screening instead of single-model extraction
No single LLM was reliable enough on its own. Having four models independently screen studies with expert arbitration caught errors that any individual model would miss.
Human-in-the-loop architecture over fully automated pipeline
Farm management advice directly affects livelihoods. Expert validation at each stage was non-negotiable for producing trustworthy recommendations.
Result & Impact
Demonstrated that LLM-assisted systematic review can accelerate the translation of agricultural research into practical farm management guidance, with findings published in Scientific Reports.
Learnings
- Multi-model consensus is more robust than relying on a single LLM for high-stakes information extraction.
- Gemini's Deep Research tool produced stronger general recommendations, but structured extraction tasks favored GPT-4.1-mini.