DagatKalat: Coastal plastic audit with photos, not servers

Problem Statement

Coastal cleanups collect bags but produce no usable data: what plastic, from what source, getting worse or better? DENR and LGUs need evidence for ordinances, but audits need experts and funding. Youth want to help but get one-off volunteer certificates, not leverage.

Proposed Solution

Transect walk + photos: the app detects and tallies sachets, PET, caps, nets, and “others” on-device, stamps a transect ID (GPS only if the team consents), and exports a CSV + photo sheet. Monthly index per sitio shows trend. Method card is one page so any school can replicate. No account server; barangay holds its own audit file.

Target Market

• Primary users: Youth eco teams, SK, coastal barangays, dive shops.
• Deployers: MENROs, NGOs, schools doing service learning.

How AI Is Used

• On-device litter detector: YOLO-tiny counts categories in cluttered sand/rocks.
• Duplicate suppression: Embedding match avoids double-counting the same bottle across frames.
• Index writer: Small LM turns tallies into “Sachet 62% — sari-sari packs dominate” lines for ordinances.

Independent Deployment Angle

• No cloud vision API, no carrier, no venue, no headset. Phones + gloves suffice.
• Open method and CSV any council can inspect; no black-box dashboard.
• Chapters compare sitio-to-sitio without a national gatekeeper.

Demo Scenario (3 minutes for judges)

1. Show a mock shoreline tarp with planted plastics; photograph with phone offline.
2. Tallies appear: sachet 14, PET 5, caps 7, net 1.
3. Second photo from another angle merges without double-count.
4. Monthly index chart shows sachet share rising — ordinance talking point.
5. Close: “Bilang bago linis.”

30-Hour Build Plan

Suggested stack: Flutter + on-device YOLO-tiny litter detector, CSV export, static dashboard. Public data / APIs: Open litter datasets + team-staged photos.

Risks & Privacy Notes

• People in photos: On-device blur for faces; audits avoid homes. Demo area contains no people or addresses.
• Overclaim: Counts are estimates with confidence; method card states limits. No shaming of named stores in v1.
• Privacy: GPS off by default; transect IDs are codes, not households.

AI Use Disclosure (draft)

Uses an on-device detector plus dedup and a small LM for index notes. AI coding assistants helped with training scripts. All demo counts were human-verified against the staged tarp.

Rubric Scoring