https://github.com/dmunish/reach
AI-powered disaster alert system for Pakistan that automatically processes official emergency warning documents.
https://github.com/dmunish/reach
ai data-aggregation disaster-preparedness disaster-risk-reduction early-warning-systems vlm
Last synced: about 1 month ago
JSON representation
AI-powered disaster alert system for Pakistan that automatically processes official emergency warning documents.
- Host: GitHub
- URL: https://github.com/dmunish/reach
- Owner: dmunish
- License: mit
- Created: 2026-01-04T05:11:57.000Z (2 months ago)
- Default Branch: main
- Last Pushed: 2026-01-27T22:09:14.000Z (about 2 months ago)
- Last Synced: 2026-01-28T07:33:59.036Z (about 2 months ago)
- Topics: ai, data-aggregation, disaster-preparedness, disaster-risk-reduction, early-warning-systems, vlm
- Language: TypeScript
- Homepage: https://reach-prototype.netlify.app/
- Size: 3.84 MB
- Stars: 1
- Watchers: 0
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
- fucking-awesome-readme - dmunish/reach - Clear visual hierarchy. Makes extensive use of HTML and custom SVG icons. Makes use of ample imagery including banner, architecture diagram and screenshots. Explains the project's purpose, implementation and roadmap. (Examples)
- awesome-readme - dmunish/reach - Clear visual hierarchy. Makes extensive use of HTML and custom SVG icons. Makes use of ample imagery including banner, architecture diagram and screenshots. Explains the project's purpose, implementation and roadmap. (Examples)
README
# REACH
Real-time Emergency Alert Collection Hub
An AI-powered warning system to bridge official disaster forecasts and the community in Pakistan.
## 
The Problem
Pakistan faces a critical disconnect in its disaster management infrastructure. While agencies like NDMA and PMD generate vital data, the "last mile" of communication is broken.
Fragmentation
Critical alerts are scattered across isolated agencies, or locked inside static PDF bulletins.
Latency & Jargon
Reports are often verbose, technical, and require manual parsing, leading to dangerous delays in dissemination.
Zero Targeting
Warnings are broadcast at the national level, causing alert fatigue. Faulty systems to deliver geofenced alerts.
## 
The Solution
REACH is an automated pipeline that ingests raw government data and transforms it into precision-targeted, actionable alerts. We treat disaster alerts as **spatial data problems**, not just text problems.
### How it works
1. **Ingestion:** Scrapers check bulletins (NDMA, NEOC, PMD) every 10 minutes for updates.
2. **Normalization:** AI processes fetched documents in under 30 seconds to extract severity, timeline, description, etc. and convert it to a CAP (Common Alerting Protocol)-inspired schema.
3. **Geocoding:** A custom service resolves location names to polygons. It handles complex directional variants (e.g., "North Khyber Pakhtunkhwa") using grid intersection logic over administrative boundaries.
4. **Distribution:** Normalized data is stored in our database and served via web app for visualization and filtering.
## 
Tech Stack
Our architecture is built for speed, resilience, and geospatial accuracy.
| **Component** | **Technology** | **Description** |
| :----------------------- | :--------------------------------------------------------------------------------------------------------------------- | :---------------------------------------------------------------------------------- |
| **Frontend** | 
| React, TypeScript, and Mapbox for web app. |
| **Services** | 
| Python microservices handling business logic and scraping. |
| **Backend and Database** | 
| **Supabase** for storing alerts, geometries, cron jobs and message queues |
| **AI Engine** | 
| Gemini-3-Flash for high-speed inference for document parsing and entity extraction. |
## 
Visuals
| **Alert Polygon and Centroid Visualization** | **Searching Historical Alerts** |
| :-----------------------------------------------------------------------------------------: | :-----------------------------------------------------------------------------------------: |
| 
| 
|
| **Filtering By Location** | **User Guide** |
| 
| 
|
## 
Roadmap
- [x] **Scrapers:** Automated bots hitting NDMA, NEOC, and PMD public sources on a 10-minute cron
- [x] **AI Pipeline:** Document processing pipeline achieving <30s latency per report using Gemini 3 Flash
- [x] **Spatial Engine:** Heuristic geocoder capable of parsing admin regions and directional descriptors into polygons
- [x] **Web Dashboard:** A responsive React application for searching alerts, filtering by severity/date, and visualizing risk zones on an interactive map
- [ ] **Deduplication:** Logic to merge overlapping reports from different agencies into a single "Source of Truth" event
- [ ] **UX Polish:** Refining the dashboard based on early user feedback
- [ ] **Alerts:** Notifications for user apps based on their GPS location
- [ ] **Performance:** Optimizing database and backend for better performance
- [ ] **Mobile Apps:** Apps for Android and iOS to get information to all users conveniently
- [ ] **Advanced Geocoding:** Improving the heuristic engine to resolve roadways, hydrology (rivers/dams), and bridges
- [ ] **Data Expansion:** Integrating social media firehose (validated) and international weather APIs
## 
Acknowledgements
- NDMA, NEOC, and PMD for their tireless work in disaster monitoring
- The open-source community for incredible tools and libraries
- Render, Modal, Supabase and Netlify for allowing us to host our app's services for free
- Communities affected by the 2025 floods - this is for you
And the [](https://github.com/matiassingers/awesome-readme) project for featuring us.
Built with 
for a safer Pakistan.