Speakers
Description
Agentic AI systems (large language models equipped with tools to read files, execute code and query databases) offer an alternative approach to beamline data analysis. Users can instruct an AI agent in natural language, and the agent in turn plans and executes the full pipeline autonomously, reducing time from data collection to interpretation.
We present a pilot project for AI beamline assistants built at ESRF from open-source components: Streamlit or Chainlit (web chat), LiteLLM (multi-provider LLM router supporting commercial and local models) and FastMCP (Model Context Protocol server). The assistant can perform custom interactive data analysis, generate visualisations and reports, or submit Slurm jobs for intensive tasks. It runs inside an Apptainer container for sandboxing, and connects to MCP servers for access to external tools or databases (for example PDB or COD). Domain-specific Skills (packaged prompt and tool-chain workflows) will be tailored for each beamline, or even for individual experiments.
We also provide an OpenCode/VSCode environment accessible from the browser. This gives scientists an AI coding assistant with direct access to the beamline Python stack, for rapid development of acquisition and analysis scripts.
