Built within a World Bank & NASA-backed RCT: a global alert system targeting air pollution advocates from 30+ cities using X (Twitter), NLP, graph analysis, and real-time satellite pollution data. Reached 200,000+ impressions in three months.
Context
Satellite instruments measure air quality almost everywhere on earth, but that data rarely reaches the people positioned to act on it. This system was built inside a World Bank and NASA-backed randomized controlled trial testing whether putting pollution data directly in front of local environmental advocates changes governance outcomes.
Approach
- Identified and ranked air pollution advocates across 30+ cities using NLP and graph analysis over X (Twitter) networks
- Ingested real-time satellite pollution readings and joined them against city-level advocate audiences
- Raised LLM classification accuracy from 65% to 87% through prompt engineering — building annotated datasets and iterative evaluation loops rather than one-shot prompt guesses
Outcome
- Over 200,000 impressions in the first three months
- A social-media scraper extracting roughly 11 million posts daily, cutting $50,000 per month in data costs
Stack
PythonNLPAWSGraph AnalysisReal-time Data