September 25-28, 2026 · Online
Free · $2,500 in prizes
Thirty-five hackathons in, across 85 countries, we know exactly what a submission and judging platform should do. So does every organizer who has ever run one. What none of us has is a modern, open, self-hostable platform that does it. This is the hackathon where you build it. The winning project is the one we run.
hackathons Hackathon Raptors has run since 2023, across 85+ countries. The spec for this event comes out of that.[15]
major hackathon platforms publish an official public API. The whole ecosystem runs on scrapers and CSV export.[14]
for one judge to score 30 projects, by the largest platform's own estimate.[1]
to build the thing that replaces all of it.
Registration, teams, submissions, eligibility, judge assignment, scoring, normalization, results, certificates, and an archive somebody can query two years later. Ten stages, each with its own state, each feeding the next. Get one wrong and the part that suffers is the judging, which is the part participants actually came for.
Hackathon Raptors has run thirty-five events on this pipeline since 2023, across 85 countries and roughly a dozen a year. That is where the specification for this hackathon comes from. We know which stage breaks first, which one nobody budgets for, and which one every organizer ends up doing by hand on the Sunday night. We have opinions earned the expensive way.
The obvious question is why nobody just uses the platforms that already exist. We studied them closely, and the answer is more interesting than "they cost money."
The incumbents converged. Devpost, Devfolio, TAIKAI, DoraHacks, HackerEarth and Unstop all ship the same nine things: an event microsite, registration, team formation, project submission, a public gallery, judge scoring, community voting, an organizer dashboard, and a CSV export.[2][6][9][11] That list is not exotic. It is a weekend of work for a team that knows what it is doing.
Then they stopped. The largest platform in the category cannot weight its judging criteria differently. Its own documentation tells organizers who need weighted criteria to judge offline in a spreadsheet.[1] Devfolio advertises "automatic score normalization" as a headline feature and, like everyone else, publishes nothing about how it works.[6][7] Community voting is universally conceded to be gameable: the standard advice is to keep the prize small and hide results until you have manually reviewed the votes.[1] And not one of them has an official public API, so the integration layer of an entire industry is unofficial scrapers and downloaded spreadsheets.[14]
Build the platform that will judge you.
[ PULL QUOTE / DEGRADED · 1-BIT ]
Meanwhile the open-source tier proves the appetite is real. Gavel came out of HackMIT with a genuine idea inside it: stop asking judges for absolute scores and ask them which of two projects is better, then recover a global ranking with a Bradley-Terry estimator.[12] JunctionApp, Dribdat, Quill and Hibiscus all exist and are all self-hostable.[13] What nobody has assembled is a modern, self-hostable, API-first whole that a working organizer can run on Monday.
That is the gap. It is real, it is narrow, and it fits in 72 hours.
Dogfood is a 72-hour hackathon with one product in it. Everyone builds the same thing against the same spec: a submission and judging portal. We publish the spec, a real anonymised dataset, and an acceptance suite you run against your own build. You ship it open source. We take the winner, self-host it, and run our events on it.
We are not asking you to build a demo of a platform. We are asking you to build ours.
The part every other hackathon leaves vague, stated plainly.
Ship under MIT or Apache-2.0. The repo is yours. We do not ask for an assignment, a transfer, a CLA, or an exclusivity clause. There is nothing to sign.
The winning project gets forked, self-hosted, and put into production for Hackathon Raptors events. That is the prize behind the prize.
Not a thank-you tweet. A credit line on every event page the platform powers, for as long as it powers them.
Every fix, hardening pass and feature we add on top gets sent back to your repository as a pull request. If you want them, take them. If you have moved on, the fork carries on and the credit stays.
We run roughly a dozen events a year and we intend to run the next decade of them on whatever wins. We are not shopping for inspiration. We are commissioning software, in the open, and paying for it.
/// One honest caveat, because you deserve it before you spend a weekend: if the top entries are close, we may adopt one and borrow ideas from another, with credit to both. We will say so publicly and in detail. What we will not do is quietly take your architecture and call it ours.
No tracks. One product, four tiers. Every team builds against the same published spec, and your score is how far up the ladder you climbed and how cleanly you did it.
Required. A submission that does not clear T1 is not judged. This is the floor, not the target.
Reference: the submission field set is stable across every platform we studied. Name, tagline, long description, thumbnail, image gallery, hosted demo video URL, repository URL, live link, tech tags, track, plus organizer-defined custom questions.[4][5]
Where the real engineering starts.
| Actor | Own scores | Peer scores | Other track | Aggregate | Audit log |
|---|---|---|---|---|---|
| VISITOR | ✗ | ✗ | ✗ | ✗ | ✗ |
| PARTICIPANT | ✗ | ✗ | ✗ | ✗ | ✗ |
| JUDGE | + | ✗ | ✗ | ✗ | ✗ |
| ORGANIZER | + | + | + | + | + |
| ADMIN | + | + | + | + | + |
Climb honestly. A clean, correct T2 scores above a broken T4 every time. Judges are reading for what works, not what is listed in your README.
The one rule that makes this judgeable, and it is verifiable in about a minute.
docker compose up produces a working portal. On localhost. Seeded with the fixture data we publish at kickoff. No cloud account, no API key, no external service, no signup.
The acceptance suite runs against your build and prints a tier-by-tier pass report. Publish that output in your repository. A judge should be able to confirm which tier you reached without reading a line of your code.
If it does not come up on a laptop with the network off, we cannot adopt it, and adoption is the entire point of the event.
After 72 hours we want software an organizer could run. Not a design system. Not a slide deck.
A working portal. It starts with one command, seeds itself, and takes a project from submission through judging to published results.
An acceptance report. The output of our suite against your build, committed to the repository. This is your tier claim and your receipt at the same time.
Honest documentation. ARCHITECTURE.md for how it fits together, DATA-MODEL.md for the schema and how an organizer gets data in and out, JUDGING.md for your scoring maths. These are read by judges and they carry weight.
What a serious submission looks like on disk.
your-portal/ ├── README.md ← what it does, how to run, honest limits ├── ARCHITECTURE.md ← the shape of the system and why ├── DATA-MODEL.md ← schema, import and export paths ├── JUDGING.md ← assignment, scoring maths, normalization ├── docker-compose.yml ← one command to a seeded, running portal ├── src/ ← your code, all of it written this weekend ├── tests/ ← your own tests, beyond the acceptance suite ├── acceptance-report.txt ← our suite's output, tier by tier ├── LICENSE ← MIT or Apache-2.0 └── .dogfood.toml ← tiers claimed, one-line pitch
One command runs it. docker compose up, seeded, working. If a judge has to read your CI config to figure out how to start it, you have failed this rule.
The acceptance report is the receipt. Claim T3 in your README and pass T2 in the report, and you score T2 with a note about the gap.
JUDGING.md counts. It feeds straight into Judging Integrity (25%). "We averaged the scores" is an answer, and it is a weak one. Tell us what you did about the judge who marks everything a 3.
Honest tier claims beat inflated ones. A team that says "we reached T2, here is the T3 work we started and did not finish" scores above a team claiming T4 with three broken endpoints.
Layout is advisory. Judges read what you actually ship.
Each project is rated on a 5-point scale across four weighted criteria. Final ranking is the weighted average across all judges who evaluated the project.
How far up the ladder you got, verified by the acceptance suite rather than by your README. T1 is a gate, not a score. Above that, correctness beats breadth: a clean T2 outranks a T4 with three features that half-work. Honest gap reporting is rewarded, inflated claims are penalised here.
The hard part, weighted like it. Is role isolation enforced in the backend or painted on the frontend? Is your normalization method documented and defensible, or did you average the scores and hope? Is there an audit trail an organizer can actually read? Did you think about vote abuse before a judge asked?
Could we run this on Monday? One command to running. Seeded with real data. Documentation a stranger can follow without asking you a question. A migration path in and out, because a platform you cannot leave is a trap. A clean license with no surprises.
Does it read as idiomatic to a senior reviewer in your stack? Is the schema one a database person would defend? And the part we cannot specify in advance: the decision that made a judge stop and say they would steal it.
Optional. Pick one and nail it. Do not half-do all four.
| Challenge | Difficulty | Points |
|---|---|---|
| Normalization Proof | Hard | +5 |
| Pairwise Mode | Hard | +5 |
| Threat Model | Medium | +3 |
| API First | Medium | +3 |
Implement cross-judge score normalization and prove it works on the fixture data. Show the raw scores, the normalized scores, and the ranking change. Document the method well enough that a statistician would not wince. Every commercial platform claims to do this and none of them will tell you how.[6][7]
Ship pairwise comparison as an alternative judging mode: show a judge two projects, ask which is better, recover a global ranking with a Bradley-Terry style estimator. This is the Gavel approach, and it sidesteps cross-judge calibration entirely by never asking for an absolute score.[12] Hard to get right, extremely satisfying when it works.
A written, defensible threat model for voting and submission abuse. Sybil votes, ballot stuffing, submission scraping, judge collusion, deadline gaming. Name the attacks you stopped, and name the ones you did not. The honest list is worth more than the heroic one.
Every action available in the UI is available through a documented API, with a published OpenAPI spec. Nobody else in this category has one.[14] Be the first.
Save yourself the trouble. These will not score well.
We are not against using a framework. We are against handing us something we cannot run.
Registration opens. Join the Discord, start sketching your schema.
Judging panel announced.
Team formation. 1-4 people per team. Solo welcome.
Raptors Conference, online and free. Unrelated event, same community, worth your afternoon.
spec.md published. Read it before the clock starts. No code yet.
Kickoff. fixtures.json and the acceptance suite released. Hacking begins.
Code freeze. Submissions due. Acceptance reports verified.
Judging window. Each project reviewed independently by multiple judges on structured forms across the window. Weighted scores and written feedback to every team.
The Teardown closes.
Winners announced. Adoption decision announced with them.
Grand Prize, and the one we run. The portal that cleared the ladder honestly, enforced its own rules in the backend, started with one command, and read like software somebody intends to maintain. This is the project Hackathon Raptors forks and puts into production.
Runner-Up. Exceptional work across the board. Strong tier completion, defensible judging maths, documentation that respects the reader. Close enough that we will be reading it for ideas.
Third Place. A standout, either for how far it climbed in 72 hours or for one decision nobody else made.
For the team whose judging layer was the most defensible. Assignment strategy, normalization method, role isolation, audit trail. The category prize for the problem the whole industry quietly avoids.
Side quest. $100 x 3.
Building a platform in 72 hours is hard. Explaining what actually happened is rarer, and more useful to everyone else. So we are paying for it.
Optional. Does not affect your main score.
What counts as a valid submission.
If we cannot run it, we cannot adopt it. Every rule below is downstream of that one.
MIT or Apache-2.0 preferred. Public at submission. You keep ownership of your repository. We do not ask for an assignment, a transfer, or a CLA.
docker compose up brings up a working, seeded portal on a laptop. No cloud account, no hosted service, no external API. If we cannot run it, we cannot adopt it.
T1 is the floor. Auth, roles, event, submission, gallery. A project that does not reach it is not scored, however good the parts are.
All project code written during the 72-hour window. Frameworks, libraries, boilerplate generators and AI assistance are all fair game. A pre-existing project of yours, or an existing open-source platform with the name changed, is not.
It runs offline on a laptop. This is not an aesthetic preference, it is the condition for the winner being adoptable at all.
Declare the tiers you reached in .dogfood.toml. The acceptance report is the receipt. Overclaiming costs more than the tier was worth.
1-4 people. Solo entries welcome, though this brief rewards a pair. Find teammates on the Hackathon Raptors Discord before or during the event.
GitHub repository, public at submission. Anonymous-username submissions accepted, but the team must be reachable for written follow-up by judges during the evaluation window.
Claude Code, Cursor, Aider, Copilot, local models, bring whatever you have. We do not gatekeep on whether you used AI. We gatekeep on whether the thing holds up and whether somebody on the team can defend the schema in writing. The acceptance report, ARCHITECTURE.md and JUDGING.md are the receipts.
This is a whole product in a weekend: auth, roles, forms, gallery, dashboards, exports. If you like shipping the entire thing yourself, the tier ladder is built for you.
The interesting problems are all yours: role isolation that survives a curl, assignment algorithms, an export path that does not lose data, an API nobody else in this category bothered to build.
The judge console and the public gallery are the product. Thirty projects to score in five hours is a UX problem before it is anything else.[1] Make that not miserable and you have done something the incumbents have not.
Cross-judge normalization and pairwise ranking are the genuinely hard parts, and the two Hard bonuses live here. Bradley-Terry, Crowd-BT, or something better you can defend.[12]
Role isolation, vote abuse, sybil resistance, audit trails. Every platform in this space treats community voting as a known-gameable feature they mitigate with policy rather than engineering.[1] Do better.
"One command to running" is not a nice-to-have here, it is 20% of the score and the reason the winner gets adopted. If you have opinions about how software should ship, this brief is downstream of all of them.
Full panel announced September 4, 2026. Reach out at hello@raptors.dev to nominate someone, or to be nominated.
Senior engineers, architects and technical leaders who have built and operated evaluation systems, run production software at scale, and sat on the other side of a judging form knowing it could have been better.
1-4 people. Solo welcome. For this brief we would recommend 2-3, since the tier ladder splits cleanly across a frontend and a backend.
Yes, and you should. Claude Code, Cursor, Aider, Copilot, all expected. We do not score whether you used AI. We score whether the portal runs, whether role isolation survives a curl, and whether somebody can explain the schema. The acceptance report and your docs are the receipts.
Yes. All of it. Rails, Django, Laravel, Next.js, Phoenix, whatever gets you to a working product fastest. The constraint here is one-command self-hosting, not dependency count. That was a different hackathon.
Then you are scored on T2 and you can absolutely still win. Tier Completion is 40% and correctness beats breadth inside it. A clean, correct, well-documented T2 has beaten a broken T4 every time we have modelled this.
Yes, inside Code Quality and inside Adoptability. A judge console that makes 30 reviews bearable is a real contribution. But a beautiful frontend over hardcoded data scores zero on Tier Completion, which is 40%.
A synthetic dataset built to the shape of a real Raptors event: roughly 40 projects, 30 judges, 8 tracks, and a full set of scores. No real names, no real submissions. It is deliberately built with the edge cases every evaluation system has to survive, including a reviewer who rates everything the same, an incomplete batch, and a duplicate entry. If your portal only works on tidy input, you will find out on Friday rather than on Monday.
A test suite we publish at kickoff that runs against your running portal and reports pass or fail per tier requirement. You run it yourself, as often as you like, and commit the output. Judges run the same suite. Nobody is guessing.
No project code. Planning, sketching schemas, reading the spec, choosing your stack, tuning prompts, all fine and encouraged. Any project code committed before kickoff disqualifies the submission.
docker compose up is the default and the safest choice. If your stack has a genuinely equivalent single command that needs nothing but a laptop, use it and say so clearly in your README. The test is whether a stranger gets a running portal from one command, not which tool printed the logs.
Any OSI-approved license. MIT or Apache-2.0 preferred, because those are the ones we can adopt without a legal conversation. Copyleft is allowed and will not cost you points.
Fork it, self-host it, and run our events on it. Credit your team on every event page it powers. Send our changes back to you as pull requests. See What happens to the winner for the full statement.
Yes, entirely. No assignment, no transfer, no CLA, nothing to sign. You are licensing your work to the world under an open license, and we are one of the people using it.
Then we will say so publicly, adopt one, and credit both, describing exactly what we took from where. We will not quietly merge someone's architecture without naming them.
Every hackathon platform in the world converged on the same nine features, and then stopped moving.
The category leader still cannot weight judging criteria and tells organizers to use a spreadsheet instead.[1] Score normalization, the one genuinely hard statistical problem in the whole domain, is advertised as a feature by multiple platforms and documented by none of them.[6][7] Community voting is treated as an unsolvable abuse surface to be managed with small prizes and hidden results rather than fixed with engineering.[1] Pricing is sales-gated across the board, so a volunteer organizer running a free event cannot even find out what it costs. And after fifteen years and millions of developers, not one of them ships a public API, leaving an entire industry to integrate through screen scrapers and downloaded spreadsheets.[14]
The open-source alternatives are better than their reputation. Gavel brought a real idea from mathematical psychology into judging and has been quietly producing fairer rankings at HackMIT for years.[12] JunctionApp, Dribdat, Quill and Hibiscus are all real and all self-hostable.[13] What is missing is a modern one, assembled as a whole, that a working organizer can deploy without becoming its maintainer.
We are not writing this as analysts. We are writing it as operators, from thirty-five events of knowing precisely which stage of the pipeline the tooling abandons you at.
Generating a CRUD app is trivial now. Building an evaluation system that is fair, that enforces its own rules, that an organizer can actually operate, and that you can hand to someone else without a handover call, is the part that still takes engineers.
That is the hackathon. 72 hours. One product. We run the winner.
[ Go to Discord ] >>>Build the platform that will judge you.