AI-Powered Multi-Agent System for Enhanced Data Collection and Processing in Invasive Species Management

Abstract

Invasive aquatic species are a problem for Europe: they are drivers of biodiversity loss and cost millions of euros each year. Europe has started a project to manage these invasive aquatic species, which begins with compiling what is already known about a species. That first step is slow. The information is scattered across separate databases and a large body of literature, in which terminology is not harmonised and stored records are often in incompatible formats. Thus building a usable picture by hand can take days, and in that time an invader keeps spreading. To address this, we developed G.I.A.S., a multi-agent LLM system that gathers, structures, and reports species data for invasive species manage- ment, built within the Horizon Europe GuardIAS project. The system uses deterministic code for the structured, rule-bound work and large language models for restructuring the non-normalised data. Each step along this tool has multiple specialised agents that assist the data flow and ease the pro- cess of compiling data about an invasive aquatic species. All data collected stays traceable and is compiled in collaboration with a researcher, ensuring an AI-human workflow. We evaluated G.I.A.S. with domain experts across two development ver- sions. Domain experts reported they would likely adopt the tool, and that it significantly decreased the time needed to produce a report. The original goal of making a report under 1 hour proved too ambitious, and the tool does not meet that requirement yet. This thesis shows that a multi-agent LLM system can be used in a scientific setting like invasive species manage- ment when it is built to stay transparent and keep the researcher in control, and that the central challenge in building a tool like this is turning chaotic, multi-source data into a structure the system can use. G.I.A.S. is released as open-source software.