What it does
Type what you're throwing away, in Korean or English. Bunri tells you which bin it goes in under Korea's household sorting rules and shows the official sentence the answer comes from, with the page and a link. If the answer depends on something (an oily pizza box, a clear or coloured PET bottle, an umbrella you can or can't take apart), it asks one question; words already in your query can answer it. If it isn't sure which item you mean, it shows three options. If the item isn't in the official sources, it says so instead of making up a rule. When a newer official rule replaces the old guide (foil-lined cartons now go with paper cartons), it shows the new rule, marks it as updated, and keeps the old sentence visible. It runs offline as a PWA and keeps an on-device pilot log so a building can measure it.
How I built it
- Turned the Ministry of Environment sorting guide into tables with the verbatim Korean text and page numbers: 285 entries from its item dictionary (pp.27–34) and 27 from its body and appendix notes, plus 3 from Seoul's 2024 page and 3 from the Korea Waste Association's 2026 Q&A (two of them override the older guide). English names, aliases and follow-up questions on top; English text is an unofficial translation.
- Matching: Korean text is decomposed into jamo so typos and spacing still match; TF-IDF over character n-grams and word tokens; a word-coverage check that lowers the score when a typed word is not explained by the matched name.
- Answer / ask / abstain thresholds are tuned on a dev split of the first query set.
- Results on a second set of 129 synthetic queries written for this project and never used for tuning: right item answered directly 36%, or among the 3 options shown 57%, so 93% resolved; 2 confident answers in the wrong bin (4.3% of answers). A substring search over the same names and aliases: 66% resolved, 24 wrong-bin answers (23%). Bunri asks more often than it answers, and in return rarely picks the wrong bin. Details and every error are in eval/results.json.
What's next
Per-district rule files, the 2026 directive table, a photo mode that feeds the same answer/ask/abstain flow, and a 4-week pilot in one dorm or apartment block.
Theme: the waste problem and how I'd measure it
Bunri targets contamination of household recycling in Korea. Paper cartons alone: of 75,000 tonnes shipped in 2022, about 60% were mixed into waste paper and 27% went into garbage bags (Hankook Ilbo, 2024). The problem is items that look recyclable but aren't (receipts, eggshells, ice packs, oily pizza boxes) going into recycling streams, and recyclables like paper cartons going into the paper pile or the trash because the rules are hard to look up. I'd measure it with a 4-week pilot in one dorm or apartment block: two baseline weeks, then two weeks with a QR code to Bunri at the recycling area. The main metric is the share of non-recyclable items in a fixed sample of recycling bags, counted on the same day each week; the on-device pilot log adds lookups, items not in the guide, and "still not sure" taps, which tell me what to add next.
Built With
- html/css
- javascript
- node.js
- playwright
- pwa
- service-worker
