A detailed account of building a scalable hackathon judging system for HackMIT, handling 1000+ participants. The old Thurstone-model-based algorithm was too slow for large events (estimated 3 years to solve a 300-team case), so the author implemented a new system based on the Crowd-BT algorithm using the Bradley-Terry model. The system uses pairwise comparisons with intelligent judge dispatch that maximizes information gain, runs as a web app with real-time rankings, and handled 1351 votes from 108 judges across 195 projects. The two-phase process included an expo judging round to select top 10 finalists, followed by panel judging for final placement.
Sort: