SagipAI: Turning national typhoon bulletins into 'what my barangay should do now'

Problem Statement

The Philippines sees about 20 tropical cyclones enter its area of responsibility every year, and warnings still get lost between the national agency and the household. PAGASA bulletins and NDRRMC advisories are written in technical English (“Signal No. 3,” “storm surge of 2–3 m”) for whole provinces. A fisherman’s family in a coastal purok needs to know whether their street floods and where their evacuation center is. During Typhoon Haiyan (Yolanda) in 2013, many survivors said they did not understand what “storm surge” meant. Barangay DRRM officers, often volunteers, also get flooded with hundreds of unstructured SOS texts and posts during a typhoon and have no tool to sort them by urgency.

Proposed Solution

SagipAI is a two-sided platform for the more than 42,000 barangays. Outbound: it takes in PAGASA tropical cyclone bulletins, rainfall warnings, and NDRRMC advisories, then overlays them on Project NOAH and HazardHunterPH flood, landslide, and storm-surge maps for each barangay and purok. An LLM writes short, specific advisories (“Purok 4 sa tabing-ilog: lumikas na bago mag-6PM papuntang San Roque Elementary School”) in the resident’s language. These go out by SMS, Messenger, and auto-generated voice calls for elderly and low-literacy residents. Inbound: residents reply by SMS or Messenger with text, voice notes, or photos. SagipAI pulls out location, the number of people, injuries, and water level, scores urgency, groups duplicate reports, and shows them on a live triage map for the Barangay DRRM Committee and the MDRRMO. Officers always approve the final dispatch.

Target Market

• Primary users: Barangay DRRM committees and officials (42,000+ barangays), plus residents of high-hazard puroks.
• Buyers (B2G): City and municipal DRRM offices (about 1,600 LGUs), provincial DRRMOs, and the OCD/NDRRMC. It can be paid for from the LGU’s 5% Local DRRM Fund under RA 10121.
• Secondary: NGOs (Philippine Red Cross, humanitarian INGOs), electric cooperatives, schools, and private employers that need staff safety alerts.

How AI Is Used

• LLM localization and personalization: Turns technical bulletins plus hazard-map overlays into short, specific actions in Tagalog, Bisaya, and Ilocano. Without it, every barangay would need a trained writer working in three languages at 2 AM during a typhoon.
• Geospatial reasoning: Joins the storm track and rainfall warnings with each purok’s hazard class to decide who gets which message. Generic province-wide blasts cause alert fatigue.
• Text-to-speech in local languages: Voice calls for elderly, PWD, and low-literacy residents who don’t read SMS.
• LLM extraction for SOS triage: Pulls structured fields (location, headcount, medical need, water depth) out of messy Taglish/Bisaya messages and groups duplicates. Officers can’t read 500 texts by hand.
• Vision model: Estimates flood depth and hazards (trapped people, livestock, rising water at roof level) from attached photos to improve urgency scores.

PLDT / Smart / ePLDT / Rokid Angle

• Smart SMS/A2P and voice APIs power both broadcast and two-way SOS. Barangay-targeted alerts on Smart’s network go beyond what the national cell-broadcast system covers.
• ePLDT cloud and data centers host LGU data locally with DPA-compliant data residency. The system can be sold as a managed service to LGUs through PLDT Enterprise’s B2G sales team.
• PLDT/Smart connectivity bundles (fiber + LTE backup + satellite failover) for barangay halls as resilient command posts.
• Rokid AR glasses (stretch): Rescue team leaders see the next triaged SOS and its route as an overlay, keeping their hands free in a boat.

Demo Scenario (3 minutes for judges)

1. Load a real PAGASA bulletin from Typhoon Kristine (2024) into the dashboard. The map highlights the riverside puroks of a sample barangay in Naga City.
2. In seconds, SagipAI generates three advisories: Bikol/Tagalog SMS for Purok 4, a Bisaya version, and a voice call. A judge’s phone receives the SMS and a phone on the table rings with the voice advisory.
3. Switch to inbound: teammates send 15 messy SOS texts and two flood photos (“tulong po, 3 bata at si lola nasa bubong”).
4. The triage board ranks them live. The rooftop family with an elderly member jumps to the top, and duplicates are merged.
5. The officer clicks “Dispatch,” and the rescue team gets the location and summary. End with the line: “From bulletin to barangay action in under 60 seconds.”

30-Hour Build Plan

Suggested stack: Next.js + Leaflet dashboard, FastAPI, PostgreSQL/PostGIS, Gemini or GPT-class LLM with JSON output, a multimodal model for photos, Google/Meta TTS, Messenger Platform API, SMS gateway. Public data / APIs: PAGASA bulletins and rainfall advisories, NDRRMC sitreps, Project NOAH hazard maps, HazardHunterPH/GeoRiskPH, PSA barangay boundaries (PSGC), OpenStreetMap (evacuation centers, roads), HDX Philippines datasets.

Risks & Privacy Notes

• Hallucinated or wrong advisories: Use templates grounded in the bulletin text, require DRRM officer approval before blasts, and always cite the source bulletin number.
• Network outages during typhoons: Queue SMS, send voice and SMS before landfall, and allow offline export to the barangay radio/PA system.
• Data Privacy Act 2012: SOS messages hold personal and sensitive data (health, location). Collect only what’s needed, get consent at opt-in, encrypt, use role-based access, and auto-purge 30 days after the event. Register a DPO with the LGU.
• Clause 24 compliance: Use only synthetic or anonymized SOS messages during development and demo. No real personal data goes into third-party AI tools. In production, use enterprise or no-training API tiers hosted in-country where possible.
• Language quality: Native-speaker review of templates and a glossary of DRRM terms.

AI Use Disclosure (draft)

We use an LLM (e.g., Gemini/GPT-class) to generate localized advisories and extract structured fields from SOS messages, a multimodal vision model to estimate flood severity from photos, and TTS to make voice advisories. AI coding assistants helped write parts of the code and draft this proposal. All AI-generated advisories and triage rankings are reviewed and approved by human DRRM officers before use, and the team reviewed all code and text.

Rubric Scoring