Archi
AI agent for CMS computing operations at CERN
Operational knowledge at CMS is scattered across wikis, JIRA, Redmine, and runbooks. Archi pulls it together — an open-source agent built at MIT that answers operator questions with cited sources and verifiable context.
Deployed at CMS
The CMS experiment operates one of the largest distributed computing infrastructures in science — thousands of physicists, hundreds of sites, millions of jobs per day. Archi is deployed as a prototype assistant for Computing Operations, answering operator questions grounded in internal documentation, JIRA tickets, and operational runbooks.
Additional instances serve SubMIT at MIT, Data Quality Monitoring shifters, and heavy-ion analysis at Fermilab.
Architecture
Archi ingests documentation, tickets, code, and web pages into a vector store, then uses a ReAct-style agent loop to retrieve relevant context and answer questions with cited sources. Tools let the agent search, fetch full documents, and query live APIs.
The system is model-agnostic (OpenAI, Anthropic, Gemini, or local models via Ollama) and ships as a single Docker deployment managed by a CLI. Interfaces include a chat app, Mattermost bot, email responder, and Grafana dashboard.