Case Studies/Faraja: Crowdsourced Crisis Damage Assessment

Humanitarian · Crisis Response Tech

Faraja: Crowdsourced Crisis Damage Assessment

Client

Independent build, inspired by UNDP's RAPIDA methodology

Location

Global

Year

2026

Services

Custom Development

UNDP's RAPIDA methodology delivers early recovery insight from satellite imagery, night-light data, and social media — but ground-level, structured field data still takes days to arrive. Kaistrum built Faraja to close that gap: an open-source, end-to-end platform that lets communities submit geolocated, AI-triaged damage reports over a mobile app or WhatsApp within 72 hours of a sudden-onset disaster, in all six official UN languages, fully offline-capable.

90.2%

AI damage-severity accuracy

Fine-tuned ResNet18 on the MEDIC benchmark (71,198 real disaster images), with 94% recall on severe damage.

500k+

Report capacity per crisis

Architecture scales from a single VPS to a 4+ replica, read-replica deployment for national-scale crises.

6

UN languages

All six official UN languages, including full RTL Arabic support, across both the web app and WhatsApp bot.

100%

Offline capable

Reports and photos queue on-device and sync automatically the moment connectivity returns.

The Challenge

Satellites see the damage. They can't see the ground.

RAPIDA fuses satellite imagery, night-light data, and social signals into a 72-hour early recovery picture — but that picture is built from orbit, and structured, ground-level confirmation remains days away due to logistical constraints. Humanitarian teams needed a way to pull community-sourced, geolocated field reports into that same 72-hour window, at a scale and reliability no existing tool offered.

  • 0172-hour assessment window — but field survey tools took days to configure and deploy
  • 02Communities span Arabic, Chinese, French, Russian, Spanish, and English — no single tool supported all six
  • 03Most-affected areas are also the least connected — low and zero connectivity by default
  • 04Existing tools broke well before 50,000 submissions — Faraja needed to hold 500,000+ per crisis, across hundreds of crises a year
  • 05Photo evidence arrived faster than any team could manually review it

The Solution

An open-source ground-truth layer for RAPIDA.

Kaistrum built Faraja as a fully open-source, seven-service platform: a mobile-first PWA and an automated WhatsApp bot for zero-barrier community intake; a Redis-backed queue that triages every photo through a purpose-trained AI model; three-tier geospatial deduplication with building-level versioning; and a live coordinator dashboard with responder dispatch — producing structured, RAPIDA-ready data within seconds of submission.

01

Reporters app + WhatsApp bot

A guided, multi-step survey delivered as a mobile PWA carousel, mirrored step-for-step by a WhatsApp bot with native location and photo sharing — no app install required, fully localised into all 6 UN languages with RTL Arabic support.

02

AI damage triage

A multi-task ResNet18 model, fine-tuned on 71,198 real disaster images from the MEDIC benchmark, classifies every submitted photo for disaster type, informativeness, humanitarian category, and damage severity — 90.2% accuracy, 20–30ms per image.

03

Three-tier deduplication & versioning

Cascading dedup — exact client-ID match, photo+coordinate proximity within 60 seconds, and 50-metre spatial matching within a 48-hour window — paired with building-level versioning so coordinators always see the latest report without losing submission history.

04

Burst-resilient queue

A BullMQ + Redis queue with append-only-file persistence absorbs submission bursts and processes reports across concurrent workers, retrying failed jobs automatically — no data lost, even through a service restart.

05

Coordinator dashboard & dispatch

A live, clustered incident map with AI-weighted zone scoring, a conversational assistant for situation summaries, and Haversine-ranked responder dispatch with optimised multi-stop routing to the field.

06

RAPIDA-ready export

PostGIS-backed spatial storage with GIST indexing, exportable on demand as GeoJSON, CSV, Shapefile, or via REST API — structured to feed directly into RAPIDA's geospatial workflows.

Outcomes

90.2%

AI damage-severity accuracy

Fine-tuned ResNet18 on the MEDIC benchmark (71,198 real disaster images), with 94% recall on severe damage.

500k+

Report capacity per crisis

Architecture scales from a single VPS to a 4+ replica, read-replica deployment for national-scale crises.

6

UN languages

All six official UN languages, including full RTL Arabic support, across both the web app and WhatsApp bot.

100%

Offline capable

Reports and photos queue on-device and sync automatically the moment connectivity returns.

Work with Kaistrum

Have a similar challenge? Let's scope it.