← All projects

The Judge

HotBuilding

Multi-agent judging for hackathons: every submission read in full, every score bound to file-level evidence.

The Judge marketing page: Every repo, read. Every score, explained., beside a spiral of contour lines.

Judging for hackathons with more teams than a panel can read in ninety minutes. An organiser picks a rubric preset and tunes the weights, teams submit a repo, and Cursor cloud agents grade it independently against the criteria — returning per-criterion scores with file-level evidence, where a score that cites none is rejected before it is stored. Each submission is pinned to a commit SHA and unpacked inside an agent's own VM, so a force-push can't change what was judged and untrusted code never runs on the server. Human judges score the same projects blind; agent and panel are blended, and every disagreement becomes a sample that calibrates the next event.

Next.jsPrismaPostgresCursor AgentsTypeScript
A score open on src/lib/route.ts, the cited lines highlighted, with an evidence stamp on the panel.
No score arrives without the file behind it. The citation is the lines themselves.
A final score of 7.7, blended from an agent score of 8.0 at 60% and a room vote of 7.2 at 40%.
The agent and the room both score. The organiser sets how much each one counts.
Three steps — write the rubric, run the Judge, weigh it in — above Bring the rubric. We read the code.
The organiser brings the rubric. The agent reads the repo.