Presenter: Michael Locher from EMPA
NestBot 2.0 is a production multi-agent AI system that gives researchers at the NEST building (EMPA Dübendorf) natural-language access to heterogeneous building data — sensor metadata, digital twin, and research outputs. This demo shows how we moved from a rigid RAG pipeline to orchestrated specialist agents: each agent has a defined tool set, runs through a central gateway with RBAC and cost tracking, and is evaluated against a structured test battery before release. No agent goes to production without a documented performance score. Intended audience: RSEs working with LLMs, research data infrastructure, or AI in science. No prior agent experience needed. Attendees will see: a live agent query against real sensor metadata, the evaluation framework that controls agent quality, and the architecture decisions that make the system reproducible and auditable.