Tantiya
Offline AI triage for typhoon damage photos, with every assessment reviewed by an officer.
- result
- Top 12 finalist at AppBuildersPH Hackathon 2026, built by a team of four. On 37 photos, two runs agreed with team-assigned damage labels 70–73% of the time. These are prototype results, not field validation.
- 2026
- ai
how it works
We built Tantiya as a team of four for AppBuildersPH Hackathon 2026 (Local AI), reaching the top 12 finalists in an event with more than 1,500 participants. The prototype helps municipal staff review photos of damaged homes after a typhoon, when connectivity may be unavailable.
After a one-time setup and model download, the app runs on one laptop: a React interface, an Express API, SQLite, and local vision and embedding models through Ollama. Staff import photos, review suggested damage categories, link photos to households, check possible duplicates, and export a report. Every assessment requires a person to confirm or change it; changes require a reason. Reports contain officer decisions, not unreviewed model suggestions.
On 37 freely licensed photos, two evaluation runs agreed with the team's draft labels on 27 and 26 photos (73% and 70%), with mean processing times of 6.9 and 6.7 seconds per photo on an RTX 4050 laptop. These are small-set prototype results, not official damage assessments or field validation. Intact nipa huts were sometimes classified as damaged; a revised prompt improved some of those cases but missed more destroyed houses. The project documents that trade-off and the need for more Philippine housing examples. Tantiya is not an official DSWD tool.
engineering details
- Top 12 finalists at AppBuildersPH Hackathon 2026, an event with 1,500+ participants; built as a team of four.
- Runs photo triage, household matching, storage, and reporting locally after setup. The repository includes an end-to-end check for requests leaving the machine.
- Requires human review for every damage assessment and records a reason when the officer changes the suggested category.
- Evaluated on 37 photos with team-assigned draft labels: 70-73% category agreement across two runs, averaging 6.7-6.9 seconds per photo on the documented demo laptop.
- Combines matching rules with local embeddings for duplicate-household checks; caught 4 of 4 planted near-duplicates with no false alarms on 3 distinct test cases.
in the product
Photo intake and review
Import field photos, inspect the model's description of the roof, walls, and structure, then confirm or correct its suggested category.
Household duplicate checks
Link photos to families and flag possible duplicate records, including similar names with spelling differences, using local embeddings and rules.
Reports from reviewed decisions
Export a CSV masterlist or printable report with counts by category and barangay, based on the officer's decisions.