GhostBaha Buster: Spot the flood project that was never built

Problem Statement

After the 2025-2026 flood-control scandal — ~10,000 projects worth ₱545B flagged as substandard, overpriced, or never built, plus arrests of lawmakers and DPWH engineers — barangays still flood every habagat while audit teams drown in PDFs. Citizens see a listed dike on paper but open water on the ground, with no way to prove it. The President ordered a full DPWH audit, but communities need their own verifier now, not after the next Typhoon Inday.

Proposed Solution

Enter a project ID or tap your barangay: GhostBaha pulls the public DPWH project list, budget, and coordinates, overlays recent satellite imagery and before/after flood maps, and asks residents for 3 geotagged photos (upstream, mid, downstream). Vision AI checks for dike, wall, or dredging signatures; an LLM writes a plain-Tagalog finding (“Nakalista: 200m floodwall 2024, ₱45M. Nakita: walang wall sa 180m”) with confidence and a printable evidence pack for COA, Ombudsman, or media.

Target Market

• Primary users: Flood-prone barangays in Bulacan, Pampanga, Cagayan, NCR esteros, youth watchdogs, campus journalists.
• B2G: LGU engineering offices, DILG field teams, COA auditors needing triage.
• Secondary: Newsrooms covering the flood-control trials at Sandiganbayan.

How AI Is Used

• Vision change detection: Compares listed structure footprint vs satellite + citizen photos; flags missing or collapsed works.
• LLM evidence writer: Turns coordinates, budgets, and image verdicts into a filing-ready Tagalog/English brief with sources.
• Duplicate clustering: Groups 50 reports of the same ghost stretch into one case so auditors aren’t spammed.
• Why AI: 10,000 rows can’t be field-checked by hand; vision + language makes citizen proof usable.

PLDT / Smart / ePLDT / Rokid Angle

• Smart: SMS short code for photo reports from non-smartphone users; alerts feed network-level flood comms.
• ePLDT: Hosts imagery cache and evidence vault in-country for COA chain-of-custody.
• PLDT Enterprise: Sells LGU “audit assist” bundle with connectivity.
• Rokid (optional): Field engineers see listed-vs-actual overlay on-site.

Demo Scenario (3 minutes for judges)

1. Search a real-style mock project: “Malabon estero wall, ₱45M, 2024.” Map shows listed alignment.
2. Satellite overlay shows open bank; upload 3 mock citizen photos — AI verdict: “NEEDS PIVOT”
3. Evidence pack generates: budget, coords, images, flood history from habagat 2026.
4. Show a legit project for contrast: “LIKELY PRESENT — 80%.”
5. Close: “Mahiya naman kayo — may resibo na.”

30-Hour Build Plan

Suggested stack: Next.js + Leaflet, FastAPI, Sentinel-2 imagery, YOLO dike-detector, Gemini LLM, Postgres. Public data / APIs: DPWH project lists, COA audit reports, Sentinel-2, Project NOAH, news archives on flood scandal.

Risks & Privacy Notes

• Defamation risk: Verdicts labeled “AI-assisted lead, needs field audit” — never accuses persons, only flags structures. Links sources.
• Data Privacy Act 2012: Citizen photos auto-strip faces/plates; no names stored without consent. Demo uses synthetic barangays and mock photos only — no real personal data into AI tools.
• IP: Satellite under open licenses; team owns prompts and schema.

AI Use Disclosure (draft)

Uses YOLO-based vision for structure checks and an LLM for Tagalog/English evidence summaries. AI coding assistants helped scaffold map UI. All demo verdicts were human-checked against mock ground truth.

Rubric Scoring