LLM-Assisted Visual Analytics for Investigating Illicit Activities

Abstract

—This work presents a visual analytics system designed to address the VAST Challenge 2025 Mini-Challenge 3. Our goal is to assist investigator Clepper Jensen in uncovering rising illegal activity in Oceanus. The project utilizes a knowledge graph derived from two weeks of radio communications, which were manually annotated by Jensen and his intern. The interface, built with JavaScript (D3.js and Graphology.js), enables intuitive exploration of communication networks between entities. Large language models (LLMs) were employed for message labeling to reduce manual investigation of messages. A pixel-based circular graph provides a rapid overview of message flows, while LLM-assisted pseudonym detection combined with heatmaps helped identify key groups: entities responsible for area protection, a group engaged in illegal operations, and a music production collective. Analysis of topic conversation peaks and communication timing exposed the use of tourism as a facade for illegal activities and preparations for a music video shoot at Nemo Reef (location). We also find evidence of Nadia Conti’s continued involvement in illicit practices. Our work demonstrates how interactive visualization and AI-driven analysis can streamline investigative workflows in complex datasets.