Constellation

Constellation is a HPC native IDE with seamless agentic AI integration that unlocks the world of HPC without needing another degree

Comment 2

About Constellation

Our inspiration was derived from the high barrier of entry to leveraging HPC for ones workload, particularly due to technical complexity. We set out to reduce this barrier. Constellation is based on the idea that parallel computing should be understood as data flow, with the technical complexity abstracted away.

Constellation is an AI-native IDE designed specifically for HPC environments. It replaces terminal-heavy workflows that require a fundamental CS education with a visual execution graph and an agentic AI interface.

So how does Constellation work?

Constellation is implemented as a web-based IDE to maximize portability and ease of use. Users upload data and describe their goal in natural language, and Constellation assists the user in constructing, executing, and debugging parallel workflows, so they can focus on their research problems.

Tasks are deployed in production clusters with AWS Batch, or ran locally on the backend for quick testing with smaller files. Scheduler scripts and batch logic are abstracted away, allowing users to focus on workload structure rather than job syntax. The AWS EC2 workers that handle these compute jobs are scaled dynamically to the workflow.

Input nodes can be constructed with single or multiple input files. Multiple input files are pipelines through the program to maximize compute efficiency, similar to a superscalar processor.

How we built Constellation

The biggest challenge we faced was constraining agent behavior so it could make meaningful changes without obscuring what was modified or why. Another major challenge was building out the cluster deployment system.

The accomplishments we were most proud of were building out actual cluster deployment rather than a simulated HPC workload, and building agentic AI integration that felt frictionless.

Our next steps include integrating with full scale HPC systems like NMC2, and using high-performance networking infra like RDMA instead of passing S3 object handles.

Our goal is to make parallel computing practical for users who do not have prior HPC experience, without constraining the capability of our users.

Built With

Share this project:

Updates