last week i joined my first ever hackathon, the appbuildersph hackathon 2026. build day was october 9, remote and overnight. demo day was october 10 at cyberzone, sm makati.
the theme was local ai. the challenge was to build an ai product that stays useful when the cloud disappears. so not an ai app for a local audience, but one where the actual ai work runs on your own device instead of someone else's server. cloud apis were allowed, but only as a side thing. the core had to work offline.
what i built
i built gel, a mac app that helps you use ai without leaking personal data.
the idea came from something i do all the time: copy a whole document and paste it into chatgpt. most leaks aren't on purpose. you just paste a payslip or a résumé without thinking about the sss number sitting in it. so gel steps in at that exact moment.
- leak guard. when you copy text with personal data and switch to an ai chat in your browser, gel warns you and gives you a redacted copy to paste instead, with placeholders like
[SSS_1]. - detection. it finds philippine government ids (sss, tin, philhealth, pag-ibig, philsys), passports, driver's licenses, bank and card numbers, gcash numbers, salaries, phone numbers, addresses, names and birthdays. it works on text pdfs, scanned pages and photos.
- file redaction. you review each finding next to the original, then save a redacted copy. the copy is rebuilt from page images, so the blacked out text can't be selected or copied back. the original stays untouched.
- ask your files. press ⌥space and ask questions in english or taglish. answers come with numbered sources pointing to the page they came from.
why local
this was the question every submission had to answer: why does this product benefit from running ai locally?
for gel it's kind of the whole point. an app that protects your personal data shouldn't send that data to a server just to check it. so everything runs on your mac:
- qwen3 4b instruct through ollama for answers
- bge-m3 through ollama for search embeddings
- apple vision for ocr on scans and photos
- apple naturallanguage for name detection
- sqlite fts5 for full-text search
- pattern matching for structured ids and numbers
the only time it touches the internet is the one-time model download during setup. there's a cloud fallback for answers, but it's off by default, and if you turn it on gel redacts the text first.
one all-nighter
building officially started at 2:30 pm, but i only started on gel around 8:30 pm. then i pulled an all-nighter and kept going until around 8 or 9 am, right before submissions closed at 10 am.
for a first hackathon, i'm happy i shipped something that works. if you want to try it or look at the code, it's on the website and on github.