Hackawon

shipitdev.github.io/hackawon

Upcoming hackathons for students, categorised by topic and sorted by what closes soonest — each one with project ideas grounded in what has actually won before. Free to use, free to run, open source.

Built for Indian students, who currently have to check three or four sites to find out what's on.

How it works

There is no server. Scraping, labelling and idea generation are scheduled work, not request-time work, so GitHub Actions does all of it and commits the results:

on a schedule, on GitHub's runners
  scrape Devfolio + Unstop + MLH        every 6 hours
    → normalise, drop finished events, de-duplicate
    → write data/hackathons/**.json     (the `data` branch is the database)
    → label each one by topic with an LLM, cached so re-runs are free
  collect winning projects from GitHub  weekly
  generate 5 ideas per hackathon        nightly, grounded in matching past winners
    → prerender a static page per hackathon, plus a sitemap
  push to the data branch → deploy → GitHub Pages serves it

Nothing runs between visits, so it costs nothing to operate and there is no server to fall over at 2am. A visitor downloads about 48 KB of data and 69 KB of JavaScript.

The data lives in this repo as plain JSON on its own data branch, so main only carries commits people made. That still gives full history through git, and a correction can arrive as a pull request that a human reads before merging.

The one exception is per-user data. Sign-in and saved hackathons go to Supabase, because they are the only things that differ per person. Listings and ideas stay static: they are identical for every visitor, so putting them in a database would add cost and fragility for nothing.

What works

Hackathons	~267 from three sources, refreshed every 6 hours
Topics	15 categories, with filters and search
Past winners	675 verified, labelled by domain, shape, and why they won
Project ideas	5 per hackathon — 1,300+, each citing the winners that inspired it
Pages	One real URL per hackathon, prerendered, readable without JavaScript
Saving	Star any hackathon; works signed out, follows you across devices when signed in
Sign-in	GitHub
Toolkit	Curated free tools and API tiers, each with a "last checked" date
Monitoring	A daily canary that files an issue when a scraper breaks or quietly returns less data

Not yet

A winners explorer · a "what tends to win" pattern library · deadline reminder emails · community-submitted hackathons · inferring Unstop's missing country field.

Running it locally

make check          # everything CI runs: pytest, ruff, tsc, vitest

# Backend — scraping, labelling, export
cd backend
uv venv && uv pip install -e ".[dev]"
../scripts/data-branch.sh pull           # copy the generated data from the data branch into data/
uv run python -m app.jobs.export_site    # build what the site reads

# Frontend
cd ../frontend
npm install
npm run dev

No API keys or accounts are needed. Labelling defaults to a fake LLM provider, sign-in hides itself when unconfigured, and saved hackathons fall back to browser storage. Tests never touch the network. To use the real services, copy frontend/.env.example and set GEMINI_API_KEY.

Layout

Path	What it is
backend/app/sources/	One adapter per site. Pure parse() + fetch(), tested against saved fixtures.
backend/app/winners/	Finds past winners on GitHub and judges whether a win claim is credible.
backend/app/taxonomy.py	Labels hackathons and projects against data/taxonomy.yml.
backend/app/health.py	Detects a source that still responds but has stopped returning data.
backend/app/ideas/	LLM providers, exemplar matching, prompt, generation.
backend/app/jobs/	The scheduled entry points.
data/	The database: JSON per hackathon and winner, plus the hand-edited taxonomy and toolkit.
frontend/src/	Static React site. data.ts is the only place app data is fetched.
frontend/scripts/prerender.mjs	Renders a static page per hackathon at build time.
docs/supabase-setup.sql	Tables and row-level security for per-user data.

Where the data comes from

Source	What we take	Notes
Unstop	Open hackathons	Under /api/public/*, which their robots.txt allows. The best source for Indian events.
Devfolio	Open hackathons, sponsor prizes	robots.txt allows all; terms contain no scraping clause.
MLH	Season events	Parsed from schema.org microdata, which they publish for search engines.
GitHub	Past winning projects	Official API.

Devpost is deliberately excluded. Its terms of service prohibit automated access, so we link out to it rather than collect from it — even though it has the largest winners dataset.

Scrapers identify themselves, stay at one request per second, and respect robots.txt. If you run one of these sites and want something changed, open an issue.

Contributing

The most useful contributions, roughly in order:

• Add a tool to the toolkit. data/tools.yml is one file and the list is explicitly incomplete. Say what you actually get, not "generous free tier", and date it.
• Fix a broken scraper. Sites redesign; the canary opens an issue when one breaks. Adapters live in backend/app/sources/ with saved responses in backend/tests/fixtures/, so you can work entirely offline.
• Add a source. Implement parse() and fetch() returning HackathonRecords.
• Improve the taxonomy. data/taxonomy.yml is a plain file. Bump version if existing labels should be recomputed.

make check must pass, and tests must run without network access.

A note on honesty

Several decisions here trade coverage for truthfulness, and changes are expected to keep that bargain:

• About half the winners have no recorded reason for winning, because most repo descriptions only say "1st place". The prompt forbids inventing one.
• "Showcased a sponsor's technology" is kept only when the text says what the project was built on. It was being inferred from event names and appeared on most labelled winners, which made it meaningless.
• Toolkit entries carry the date a human last checked them and grey out after 90 days. Free tiers change constantly — xAI's free credits ended in May 2025 and much of the web still recommends them.
• Where grounding is thin, the site says so rather than implying confidence it does not have.

Licence

Code is MIT. The collected data in data/ is CC BY 4.0. Hackathon listings and project descriptions belong to their original authors — we store facts and link back rather than republishing their content.