Workshop: HUST-26: 13th International Workshop on HPC User Support Tools
Authors: Andrew Pasquale (Sweet & Fizzy Digital), Vikram Gazula (University of Kentucky), and Joseph E. Bacal (Sweet & Fizzy Digital)
Abstract: RCD/HPC centers increasingly run support chatbots, but these answer from static documentation and cannot handle information that changes or act when a user needs something done. We describe our experience building an agent for a national cyberinfrastructure program that both answers questions from live infrastructure data and takes actions on the researcher’s behalf, such as registering them for events, posting announcements, and filing tickets, under their own federated identity, with user data kept inside institutional infrastructure and a confirmation step before consequential writes. Because help is only useful if it stays good, we evaluate answers with a judge validated against human reviewers, which separates the agent from plain retrieval (0.97 vs. 0.47), and we run that evaluation two ways: a release gate that catches our own regressions, and a drift detector that catches when the underlying information has gone stale.
Back to HUST-26: 13th International Workshop on HPC User Support Tools Archive Listing
Back to Full Workshop Archive Listing