From one idea to 19 projects: the Hi Youth hackathon’s ModelBest edition wraps up
Give yourself 22 hours, one idea and an open-source agent operating system. How far can you get?
On September 11 and 12, the Hi Youth Haidian Young Creators Bootcamp Hackathon — ModelBest Edition took place at Yuandian Academy. This time, Yuandian Academy teamed up with ModelBest and OpenBMB to put PilotDeck in the hands of 52 university students, giving every idea a chance to become a real product.
From forming teams and picking topics to writing code, building products and tuning models; from laptop screens still glowing at dawn to demos that could actually be tried and judged the next afternoon —
the students took on real needs and ran an innovation sprint that pushed ideas quickly toward real scenarios.
A rigorous selection: 52 students make the cut
More than 130 university students applied, and after several rounds of screening, 52 made it to the venue.
They came from 18 universities and research institutes, including Tsinghua University, Peking University, Beijing University of Posts and Telecommunications, Beijing Institute of Technology, The Hong Kong Polytechnic University and HKUST. Their fields ranged from computer science and AI to medical engineering, law and design. Some had just opened their first undergraduate textbook; others had spent years in a lab. Different schools, different paths, one shared trait: curiosity. Curiosity about a problem, about a piece of code, about what they could still build.
People from different majors form impromptu teams around a single idea. This is where the hackathon gets interesting.
Three tracks: understand, use and improve PilotDeck
Before the contest began, PilotDeck project lead Yan Yukun explained what an agent operating system is and walked through PilotDeck's core technical features: project-level isolation built around WorkSpace, white-box memory, intelligent routing and the Always-on mechanism.
The hackathon offered three tracks:
Track 1 · Scenario mini-tools: build AI-native applications on PilotDeck and put agent capabilities into real, small scenarios;
Track 2 · Multimodal coordination: orchestrate multimodal models to build projects that can see, hear and generate;
Track 3 · Hardcore geek: take on the underlying architecture and extend, optimize or refactor PilotDeck itself.
After choosing a track, the contestants began a full day of intense development with mentors from ModelBest.
19 projects that put AI into real, small problems
In the end, 19 projects were submitted. Rather than debate a grand "AI future," this cohort preferred to start from concrete problems:
One team was tired of the inefficient loop of generating AI images, manually rerolling, and debugging them by eye, so they built "A Thousand Images, a Hundred Refinements · AI Image Generation Supervisor" — it uses PilotDeck's Agent Loop to run a generate-review-revise workflow, giving the AI its own quality bar so it finds its own errors and redoes its own work. Another team struggled with oversized goals and the inertia of never starting, so they built "FryPlan," an AI action-starter that breaks big tasks into lightweight small goals and uses visible progress to lower the threshold for taking the first step. A third team targeted delivery quality when on-device models execute agent tasks, designing "Robin," a two-layer acceptance mechanism for subtask delivery in PilotDeck.
Fresh projects face real users right away
On the afternoon of September 12, the venue switched from hacking mode to an open project fair. All 19 projects were on display; visitors moved freely between booths, heard the teams pitch, tried the demos and talked with the developers face to face: why build it this way? What problem does it really solve? What could it become next?
Meanwhile, Track 1 and Track 2 projects entered Origin Leap, an on-site interactive session where visitors upvoted the projects they liked and the likes fed into judging — a project had to do more than get built; it had to hold up when real users tried it.
In the evening fireside session, the PilotDeck core team and contestants talked and sparred over industry trends, personal growth, research and careers.
Overall judging plus a hardcore special award: which projects came out on top?
After the judges tried every project and deliberated on site, the winners were announced:
First prize | Robin: Knowing the Mistake and Changing
Second prize | A Thousand Images, a Hundred Refinements · AI Image Generation Supervisor
Third prize | PilotYoke
Third prize | FryPlan
Third prize | Later: Not Now But Someday
Special award · Hardcore Geek | PilotRoute: an intelligent inference routing system combining model sharding, cache reuse and failure fallback
The contest ends; the connections don't
From check-in and team formation on day one to product showcases at the fair, contestants polished rough ideas into complete products that could be tried and tested. In just 22 hours they compressed the full product loop: finding a problem, co-creating in teams, iterating toward launch and collecting real user feedback.
PilotDeck also received several pull requests from contestants. Over the course of the hackathon, an open-source project was genuinely used, picked apart, improved and left better than it started.
A contest has a finish line; open-source co-building does not. PilotDeck is now open source to developers worldwide (https://github.com/OpenBMB/PilotDeck). ModelBest and OpenBMB invite every developer passionate about agent technology to download it and put it to the test — whether you wire the agent operating system into your own workflow or join the build with a pull request, as this cohort did, you become part of PilotDeck's evolution. We look forward to meeting more fellow travelers in the open-source community as we explore where the agent operating system goes next.