INK (Interactive aNalysis worKbench) is a web-based interactive computing platform developed at the IHEP Computing Center, giving physicists browser-accessible interactive sessions — Jupyter, VS Code, ROOT, VNC, and remote desktops — transparently backed by our production HTCondor and Slurm clusters. By abstracting job submission, resource allocation, connection management, and authentication behind a unified web interface, INK lets users launch and connect to interactive workloads with a single click, without writing submit files or handling HTCondor directly, while still benefiting from the scheduling and resource management of the batch system underneath.

Key features of INK include a pluggable job-type framework that makes new interactive applications easy to add, a session-based reverse proxy for secure browser access to running jobs, and a portable, site-agnostic core with IHEP-specific logic isolated as plugins.

Building on this platform, we have been exploring how AI — large language models in particular — can be delivered on HTCondor. This talk presents our integration of AI coding assistants (OpenCode, OpenChamber) as first-class interactive job types, along with the design choices for running LLM-based tools securely on shared clusters. We will share the architecture, our experience, and the practical challenges of bringing AI applications into a production HTCondor environment.

Associated Links

• Presentation Slides (PDF)
• Presentation Slides (PPTX)
• Indico Agenda