RoundOne
Hackathon Judging
Agent & Management
Collect submissions, run judging rounds and announce winners from one app you deploy for your org. The RoundOne agent, built on NVIDIA Nemotron and Nebius, checks each project’s code and demo against your rules, so your judges can spend their time on the ideas.
An entry in the Nebius × NVIDIA Global AI Hackathon ↗
RoundOneRuntime 3:24Thank you to the sponsors of this hackathon & Devpost
AI makes a bad hackathon judge
Most submissions are written with AI now, so they're tuned to hit every note in your requirements. An AI judge grades work shaped like its own, and it can give 8 lines of code a 10 out of 10.
AI grading AI
The project says all the right things, so it gets full marks.
- Uses the sponsor APIYes
- README covers every criterionYes
- README1,400 words
- Code8 lines
that wrote it
What AI is good at
It can’t judge the idea, novelty, or creativity behind a project. It can do the groundwork, so a person can.
- VerifyDoes it call your API? Does it build and pass its tests?
- ClassifyWhich projects are most alike?
- ConfirmDoes it meet the base requirements, like a public repo and a demo under three minutes?
- SummarizeA few sentences on what each project is and how it did on your checks.
RoundOne doesn’t replace human judges.
It lets them judge at scale.
Hackathons get your product into builders' hands
Give developers a deadline and a reason to build something real on your API, SDK or model.
Developers who know your product
Every team has to learn your product well enough to ship something with it. Some keep building after the event.
Feedback on your docs and API
People working to a deadline find the confusing errors and missing examples fast, and they tell you about them.
Use cases you hadn't thought of
Teams point your product at problems that weren't on your roadmap, and some of those ideas are worth a closer look.
Demos and examples
Each project comes with a repo and a demo video. With the team's permission, the best ones make good examples and tutorials for your docs.
Pick winners for their ideas
If you judge on technical features alone, a polished project built with coding agents can beat a better idea. And if the team that ran the most coding agents wins, expect more of the same at your next hackathon. RoundOne hands the technical checks to its agent and leaves the ideas to your judges. See how it works →
Coding agents multiplied the submissions
With a coding agent, a small team can ship a working app over a weekend. Your judging panel stayed the same size.
Ten times the projects for the same three judges
before anyone clones a repo
RoundOne checks the code so your judges can score the idea
A coding agent can write the app, but the idea and the taste behind it still come from the team. You choose, criterion by criterion, where the RoundOne agent helps and where it stays out.
What the RoundOne agent checks
- Does it build and pass its tests?Sandbox run
- Does it call your API or SDK?Code scraper
- Is the demo video under the time limit?Video reviewer
- Is the live demo up?Web scraper
What your judges decide
- Quality of the ideaIs it new, and is it worth doing?Human only
- Potential impactWho is it for, and how much would it help them?Human only
- DesignIs it clear, and was it made with care?Human only
Judges still score every criterion. Set one to Human only and the agent leaves it alone.
For example, on the Nebius × NVIDIA Global AI Hackathon rubric, the agent checks the sponsor rules as must-pass gates and leaves Quality of the idea, Potential impact and Design to the judges.
You and your judges pick the winners
- Ranked by your judgesProjects that pass your gates rank by your judges’ average score. The agent’s score only breaks ties.
- Gate failures come to youWhen the agent fails a project on a gate, the project waits in your inbox. Keep it out, or put it back in the pool.
- You set the final orderAfter the last round closes, you can drag projects into a different order before you publish, and you pick the award winners yourself.
- 1SGSignal Garden8.77.2
- 2TWTidewatchTie broken by agent8.76.8Tie broken by agent
- 3RWRepo WhispererAgent's highest8.19.0Agent's highest
- 4PPPantry PilotFailed a gate · kept out8.97.4Failed a gate · kept out
Sorted by the judges’ average, then the agent’s score, then submission order. Gate failures go last.
NVIDIA’s open models on Nebius let the agent check every project
Checking 1,000 projects means building 1,000 repos of untrusted code and making a lot of model calls. Nebius runs both. NVIDIA Nemotron models do the agent’s scoring, and Tavily reads the live web.
Nemotron
Out of the box, every agent score and pass/fail verdict comes from a Nemotron model. You pick the tier for each criterion.
- QuickNemotron 3.5 LightningOne pass over what the checks found.
- BalancedNemotron 3 Super 120BChecks a few things itself with tools. It also runs chat.
- In-depthNemotron 3 Ultra 550BThe largest model, with the most tool calls.
Double check. When a verdict has low confidence, raises a flag or fails a gate, the agent runs it again one tier up. In-depth criteria get a second Ultra pass. That’s the default, and you can switch it to Always or Off.
Vision models on Token Factory describe the demo video’s frames. A Nemotron model does the agent’s scoring.
Token Factory
Token Factory serves the model calls and the embeddings behind chat search. Each repo is cloned into a VM-isolated Nebius Sandbox, and each review step works on its own fork, so one step’s installs never touch another’s.
- Extract reads a project's pages.
- Crawl walks the live demo on In-depth criteria, following your instructions.
- Search checks whether an idea already exists on the web.
Chat uses it to search the web too.
One hackathon, from rubric to winners page
Setup, judging and results in RoundOne, screen by screen. Click one to see it full size.
Setup · 01 / 17
Screen 1 of 17: Shape every submission.
Take it for a spin
Sign in to the demo and you’ll find three sample hackathons: one finished, one mid-judging and one just set up. Click anything you like. Nothing you change is saved.
A queue a judge can finish
You can release a round in daily batches, so a judge only sees today's projects. Each judge gets the projects in a different order, which keeps the same few from always landing at the tired end.
Built for the judge on project forty
- No account neededJudges open their private link and start. They see only their own queue.
- A time estimate up frontThe start screen shows how many projects are waiting, what they’ll score and about how long it will take.
- Score first, then compareA judge sees the agent’s verdict only after submitting their own scores. Gaps of 2 points or more get flagged, and the judge can change their score once.
- Scores stay privateJudges never see each other’s scores, so one judge can’t anchor the rest.
Group review · round 1 of 2
Welcome, Grace
4 of 12 done today, 8 to go. 36 more open over the coming days.
Your projects today
- #118Tidewatch7.5
- #042Repo Whisperer8.0
- #207Quiet Hours6.5
- #033Pantry Pilot7.0
- #251Signal Gardennext
- #096Loom Lens
A smaller pile each round
Set the gates a project must pass and how many projects advance. Then watch each round fill in, and close it when it's done.
- Submitted186all projects
- Screening142passed the gates
- Round 140live64% reviewed
- Finals12top 12 planned
Self-host it for your whole org
RoundOne is a Next.js app on Supabase. Everyone in your org can run hackathons on one deploy, and each organizer sees only their own.
- Your deployYou need a Supabase project and a Node host such as Vercel, plus a Nebius Token Factory key for the agent and a Nebius project ID for its code sandboxes. Tavily and email are optional.
- Your databaseSubmissions and scores stay in your own Supabase database.
- Your brandOne config file puts your name, logo and color on the whole app. Each hackathon adds its own logo and color to its submission form, judge portal and winners page.
- Your modelsSwap the model behind any tier with one environment variable.
