Tech Stack
Overview
On June 24, 2026, a M7.2 earthquake struck near Yumare, Venezuela. Within hours, official channels were overwhelmed and families had no reliable way to find damage reports, shelter locations, or missing persons information.
I built a real-time damage map the same day — aggregating reports from YouTube, X/Twitter, and Instagram every 10 minutes and plotting them on an interactive heatmap. The app reached 4,000+ people in the days following the earthquake.
What it does
- Damage heatmap — weighted by report credibility and damage severity, updated every 10 minutes
- Drill-down reports — click any zone to see the underlying videos, tweets, and posts that generated the signal
- Relief centers — verified centros de acopio with addresses and accepted items
- Missing persons — community-submitted reports plotted on the map
- Emergency directory — 30+ Caracas hospitals, ambulances, bomberos, and rescue teams
- USGS PAGER — live fatality projection pulled from official seismic data
Data pipeline
The system ran three parallel scraper processes on a local Mac Mini, each on a 10-minute interval:
- YouTube — 6 search queries, auto-subtitles with Whisper fallback for Spanish transcription
- X/Twitter — 13 Playwright searches including verified accounts (convzlacomando, MariaCorinaYA)
- Instagram — 5 hashtags via saved session
- Geo-extraction — Gemma 3 4b running locally via Ollama to extract location references from Spanish text, zero API cost
- PostGIS — geospatial queries for clustering and heatmap generation
The architecture was deliberately built to run without cloud compute costs during the acute phase — everything ran on a Mac Mini with local inference.
Why I built it
Venezuela is where I’m from. When the earthquake hit, the information fragmentation was immediate — reports scattered across WhatsApp groups, YouTube livestreams, and Twitter threads with no unified picture of where damage was worst or where relief was going.
The tools I had — Mapbox, Supabase PostGIS, Python scrapers, local LLM inference — were exactly what was needed to build a real signal out of that noise. So I built it.
This project is an example of what I think software can do at its best: respond to a real crisis with real infrastructure, fast enough to matter.
Impact
- 4,000+ page views in the days following the earthquake
- Used by families searching for missing persons and damage information
- Relief center data surfaced donation drop-off points across Caracas