ConnectPH: AI that finds the Filipinos the internet forgot, and ranks where to connect them first

Problem Statement

Millions of Filipinos, especially in mountain and island barangays classified as GIDAs, still lack reliable internet. The archipelago’s more than 7,000 islands and its rugged terrain make planning backhaul expensive. Planners often depend on outdated census counts and assumptions made at the municipal level, so the actual clusters of unserved households stay invisible. DICT’s Free WiFi for All program (RA 10929) and private rollouts face the same question: where does the next site deliver the greatest social and commercial return? Without granular evidence, budgets get spread thin, sites go up where demand is weak, and isolated schools, rural health units, and barangay halls stay offline during typhoons, when connectivity matters most.

Proposed Solution

ConnectPH Map is a geospatial AI planning tool. A building detection model, together with open footprint layers such as Google Open Buildings and Microsoft footprints, maps every rooftop from Sentinel-2 and high-resolution imagery. It fuses these with gridded population (Meta HRSL/WorldPop), PSA census data, crowdsourced coverage and speed data (Ookla open data, OpenCelliD), roads, terrain, and hazard layers. A model then estimates unserved population per 1 km² grid cell. For each cluster it recommends the cheapest viable technology: fiber extension, macro or small-cell tower, satellite/VSAT, or a free WiFi hotspot at a school or barangay hall. An LLM assistant lets planners ask in plain language, for example “Top 20 unconnected barangays in Samar with a public school and no road access,” and get back site briefs ready for export.

Target Market

• Primary users: PLDT/Smart network planning and PLDT Enterprise B2G teams.
• B2G buyers: DICT (Free WiFi for All, the National Broadband Plan), DepEd (school connectivity), DOH (rural health units), provincial and city LGUs, and NEDA/DEPDev.
• Other buyers: Development partners (World Bank, ADB, USAID-type programs), tower companies, and satellite ISPs.
• Size: All 42,000+ barangays, with a priority focus on GIDA municipalities in Mindanao, Eastern Visayas, Bicol, MIMAROPA, and the Cordillera.

How AI Is Used

• Computer vision (segmentation) detects buildings and settlements in satellite imagery, including new informal settlements that census data does not show. Without it, planners only see municipal totals.
• Machine learning regression estimates unserved population and demand (likely subscribers, ARPU proxy) per grid cell from fused features.
• Optimization and ranking (multi-criteria scoring plus greedy or ILP site selection) picks the technology and the order of sites to maximize people connected per peso, weighting schools, health units, and hazard exposure.
• LLM planning copilot turns natural-language queries into geospatial filters and writes site justification briefs for DICT/LGU proposals.
• Without AI, the same analysis requires months of manual GIS work for each province.

PLDT / Smart / ePLDT / Rokid Angle

• Directly serves PLDT/Smart’s core business: it helps prioritize fiber (PLDT has the country’s largest fiber network), towers, and Smart’s satellite and LTE/5G expansion.
• PLDT Enterprise B2G: proposals backed by evidence for DICT Free WiFi, DepEd school connectivity, and LGU smart-city deals.
• ePLDT: hosts the geospatial data lake and models on ePLDT cloud/VITRO data centers and can be sold to LGUs as a SaaS dashboard.
• Rokid AR glasses: field engineers on site surveys see candidate tower locations, line-of-sight, and terrain overlays hands-free.
• Supports ESG: it gives measurable “digital inclusion” metrics for PLDT sustainability reporting.

Demo Scenario (3 minutes for judges)

1. Open with a photo of a student in Northern Samar climbing a hill to catch signal for an online class.
2. Zoom the map to Northern Samar. Satellite tiles appear and the AI outlines thousands of rooftops in seconds.
3. Turn on the coverage layer. Red cells show dense rooftop clusters with no measured coverage.
4. Click “Rank.” The top 10 sites appear, each with a recommendation such as “Barangay X: 2,300 people, public school, 6 km from fiber → fiber extension, est. ₱/person connected.”
5. A judge types: “Show unconnected barangays with a rural health unit and high typhoon risk.” The copilot filters the map and generates a one-page DICT proposal PDF.
6. Close: “PLDT can connect 50,000 more Filipinos with the same budget, because it now knows exactly where they live.”

30-Hour Build Plan

Suggested stack: Python, GeoPandas, H3, PostGIS, PyTorch (U-Net/SAM-based segmentation), scikit-learn/XGBoost, OR-Tools, FastAPI, deck.gl/Leaflet, Gemini or an open LLM. Public data / APIs: Sentinel-2 (Copernicus), Google Open Buildings, Microsoft Building Footprints, Meta High Resolution Settlement Layer, WorldPop, PSA census (OpenSTAT), Ookla Open Data, OpenCelliD, OpenStreetMap, DepEd school locations (data.gov.ph), DICT Free WiFi site lists where public, HazardHunterPH and Project NOAH hazard layers.

Risks & Privacy Notes

• Coverage data is incomplete: crowdsourced data under-samples areas with no users. Treat the absence of tests as a signal and give confidence intervals. In a real deployment, plug in Smart internal coverage data under NDA.
• Imagery resolution: Sentinel-2 at 10 m misses small houses. Combine it with open footprint layers and validate on sample tiles.
• Data Privacy Act of 2012: only aggregated, non-personal data is used (grid-level population, anonymized speed tests). No subscriber data is used in the prototype.
• Fairness: ranking weights must not ignore low-revenue communities. Include a social-priority weight for GIDAs, schools, and health units.
• Clause 24 compliance: no personal or confidential (including PLDT internal) data goes into AI tools. AI use is disclosed, and licenses are respected (e.g., Open Buildings CC BY 4.0, OSM ODbL, Ookla CC BY-NC-SA).

AI Use Disclosure (draft)

We use PyTorch segmentation models (open source, optionally SAM) for building detection, XGBoost for demand estimation, and Gemini/an open LLM for the planning copilot and brief generation. GitHub Copilot and ChatGPT helped with boilerplate GIS code and pitch drafting. All rankings, model outputs, and generated briefs were reviewed by the team and checked against source data.

Rubric Scoring