Multi-Agent Data Visualization and Narrative Generation

Team

Anton Wolter, Georgios Vidalakis, Michael Yu, Ankit Grover, Vaishali Dhanoa

Problem

Agents could support the whole path from data exploration to communicating insights. But when every step runs through a language model, the result is hard to inspect, and a single mistake means regenerating everything.

Approach

We combine a multi-agent architecture with deterministic components and move critical logic out of the language model, to make the pipeline more transparent and reliable.

System

The system produces granular, modular outputs, so people can change one part of a narrative without regenerating the whole.

Findings

We evaluated the system on four diverse datasets, looking at generalizability, narrative quality and computational efficiency.