At the Lawrence Berkeley National Laboratory’s Advanced Light Source (ALS) particle accelerator, an AI-powered system called the Accelerator Assistant is helping keep one of the world’s most complex scientific instruments running smoothly. This large language model (LLM)-driven tool, built on an NVIDIA H100 GPU and powered by CUDA for fast inference, is designed to support high-precision X-ray experiments by streamlining operations and reducing downtime. The ALS accelerates electrons to near light speed in a 200-yard ring, producing intense beams of ultraviolet and X-ray light that travel through 40 beamlines. These beams enable about 1,700 scientific experiments annually, supporting research in materials science, biology, chemistry, physics, and environmental science. When the system goes offline, even briefly, it can disrupt dozens of experiments in progress, with recovery times ranging from minutes to days. With over 230,000 process variables in its control system, diagnosing and fixing issues has traditionally been a time-consuming, high-pressure task. Operators had to manually locate data, consult experts, and coordinate responses under tight deadlines. The Accelerator Assistant now automates much of this work, using context-aware prompts and access to institutional knowledge, historical data, and real-time system information. The system runs on a hybrid architecture. It can perform on-premises inference using Ollama and an H100 GPU within the control room, or route requests through the CBorg gateway to external models like ChatGPT, Claude, or Gemini. This ensures both speed and security. All interactions are authenticated and tied to user sessions, allowing personalized context and memory across multiple tasks. The assistant uses Osprey, a custom framework developed at Berkeley Lab, to safely apply agent-based AI in complex control environments. It integrates with EPICS, the standard control system for large scientific facilities, ensuring that all actions adhere to safety protocols. Engineers can use natural language to describe goals, and the system translates these into precise Python scripts that analyze data, generate visualizations, or safely interact with hardware via Jupyter Notebook environments. According to Thorsten Hellert, staff scientist and lead author of the research, the system can prepare and run multi-stage experiments in minutes—cutting setup time by up to 100 times. This efficiency is achieved through carefully engineered prompts that embed relevant context, such as beamline configurations, sensor locations, and operational history. The AI agent acts as a specialized expert, drawing on curated knowledge to guide its reasoning. It can locate specific sensors, retrieve past experiment logs, and suggest troubleshooting steps—all without requiring deep technical expertise from the user. Looking ahead, Hellert aims to build a comprehensive wiki of operational procedures to further empower the AI agents. This would allow the system to operate with greater autonomy, always with a human in the loop to approve critical actions—especially important for high-value equipment like microscopes or experimental setups worth millions. The framework is already being adopted across other U.S. particle accelerator facilities as part of the Department of Energy’s Genesys mission. It is also being tested for use in the ITER fusion reactor in France, the world’s largest fusion project, and in the Extremely Large Telescope in Chile. Beyond operations, the ALS’s work has had profound scientific impact. During the pandemic, it helped identify a key antibody that neutralized SARS-CoV-2, accelerating the development of effective treatments. It also played a central role in research on metal-organic frameworks, materials that capture carbon dioxide and water from air—work that contributed to the 2025 Nobel Prize in Chemistry. Additionally, ALS analyses of asteroid samples from NASA’s OSIRIS-REx mission provided evidence that early Earth may have been seeded with water and organic compounds from space. By combining AI with cutting-edge science, the Accelerator Assistant is not only keeping the ALS running but also advancing discoveries that benefit humanity.