BEACON turns community phones into a verified damage-reporting network. Reports become triaged, geo-tagged, AI-credibility-scored data points for UNDP and partners — in minutes, not days.
Watch the 2-min walkthrough 02:11Photo + GPS + optional landmark.
Camera intent via PWA share-target. GPS resolved to ±10m or landmark text accepted as fallback.
On-device CV picks a grade; EXIF + perceptual hash checked.
MobileNet-v3 quantized to 8-bit, ~14 MB, runs in browser ML Kit / TFLite. No upload until classified.
Online instantly or queued offline; cluster-detected if duplicate.
Background Sync API + Cloudflare Queue. Perceptual hashes compared against last 6h of submissions for de-duplication.
6-factor credibility score; operator reviews if needed.
EXIF geomatch · event-window correlation (USGS/EMSC) · news cross-ref · image authenticity (C2PA) · CV damage grade · duplicate detection.
Verified data flows to UNDP partners via GeoJSON / API.
Webhooks to OCHA, HOT, IFRC. GeoJSON, CSV, OData feeds. Per-org rate limiting & API keys.
Public-facing reporting wizard with 7-language support, OSM building-footprint map, on-device AI pre-classification, and offline queue.
Triage queue, 6-factor AI credibility, cluster detection, GeoJSON / CSV / OData export.
Installable, offline-first PWA. Capture, queue, sync. Works on Android Go and 2 G networks.
EXIF geomatch · event-window correlation against USGS/EMSC alerts · regional news cross-reference · C2PA + perceptual hash · CV damage classifier · cross-submission duplicate detection.
Faces and license plates are detected and redacted on-device before any upload. Reporters stay anonymous by default.
Works in zero-bar areas. Reports persist in IndexedDB and sync automatically when connectivity returns.
The six UN languages plus Burmese for the Mandalay scenario. Reporters write in any language — translation runs on-device.
When GPS is unavailable, a plain-language description is geocoded by an operator — no report is lost because of a missing pin.
When the AI score is low, BEACON requests additional proof anonymously instead of auto-rejecting. Real reports survive ambiguity.
| BEACON | Ushahidi | KoBoToolbox | WhatsApp tipline | |
|---|---|---|---|---|
| AI credibility scoring | YES · 6-factor | No | No | No |
| Offline-first with queue | YES · PWA + sync | Partial | Yes | No |
| Building footprint overlay | YES · OSM live | No | No | No |
| Auto-blur PII client-side | YES | No | No | No |
| Landmark-fallback geocoding | YES | Partial | No | Manual |
| Cluster / duplicate detection | YES · auto | Manual | Manual | Manual |
| Document-escalation flow | YES · anonymous | Manual | Manual | Manual |
| Open source | YES · MIT | AGPL | AGPL | No |
| Multilingual UI (UN 6 + local) | YES · + Burmese | Yes | Yes | n/a |
| Rapid deploy in < 48 h | YES · one-line deploy | Days | Hours | Hours |
UNDP triggers deployment via a single Terraform plan. Cloudflare Workers, R2, Postgres+PostGIS, Queue all spin up in under 8 minutes. Region pinned to nearest landing zone.
QR code and SMS shortlink distributed through local broadcasters, telco partners, and NGO networks. Web PWA installable from the QR — no app store dependency.
Community reports begin arriving. Operator console is live for the assigned coordinator pool. Auto-cluster + auto-verification triages 70% without manual review.
Burmese voice-input enabled. Satellite imagery layer for the affected region overlaid via Copernicus EMS rapid mapping feed.
WhatsApp bot enabled for users without smartphones. Cluster summaries piped to OCHA situation reports. Daily 06:00 / 18:00 sitrep cadence.
Field-team mission briefings are generated entirely from BEACON data. Building-level damage atlas exportable as GeoJSON and printable map books.
Each prototype is reachable from the cards above. The fastest way through is to walk the same scenario end-to-end: a reporter submits, an operator verifies, the data flows out.
Open the community web app. Start a damage report. Toggle offline mid-flow to see the queued state — then come back online to watch it sync.
Open the operator console and click submission BCN-26-7846 (the synthetic reverse-image case) to see the document-escalation flow rather than an auto-reject.
Open the mobile PWA and follow home → capture → submit → queue. Toggle the offline pill to see the on-device queue persist across page reloads.
BEACON is released under MIT. A non-exclusive, royalty-free, perpetual license is granted to UNDP and its partner agencies per the Innocentive challenge brief. The source repository can be transferred or mirrored to a UNDP-controlled org on award.
All examples on this page use the synthetic "Mandalay M6.8" scenario. No real victims or addresses are depicted. The dataset is generated from window.BEACON.submissions and is reproducible.