Speaker
Description
Following the start of stable data taking at the Jiangmen Underground Neutrino Observatory (JUNO), reliable link monitoring is required beyond standard data-quality checks. In particular, the Back-End Card subsystem is a key component of trigger links, and a recurring operational challenge is rapid and accurate root-cause localization when a link drops across multiple hardware and software layers. While JUNO has accumulated many Python diagnostics to inspect counters, logs, and topology-dependent symptoms, conventional diagnostic procedures are unscalable and fragmented across scripts.
This study develops an AI agent architecture that makes Large Language Model (LLM) based assistance usable and maintainable in such complex scientific software ecosystems by addressing three core problems: reliable tool execution, reliable long-task solving, and scalable onboarding of new tools. The proposed two-mode framework comprises an Analysis Agent that invokes tools through Skill Cards and maintains long-term task coherence via a Folded Memory mechanism. Skill Cards are structured, type-safe contracts capturing argument schemas, usage policies, and safety constraints. Second, to resolve the bottleneck of manual tool integration, a Learning Agent that automatically converts raw Python functions into Skill Cards and iteratively validates them to prevent schema drift, runtime failures, and tool distraction in a growing registry.
The resulting workflow orchestrates existing diagnostics with engineering-grade determinism while preserving LLM flexibility in procedure selection. Initial evaluations indicate improved consistency, stronger traceability, and better success rate compared to standard AI Agents while keeping the diagnostic library modular and maintainable as it evolves.
| Minioral | Yes |
|---|---|
| IEEE Member | No |
| Are you a student? | Yes |