GreenWatch AI

Inspiration

When people think about sustainability, the solution usually sounds simple: execute more climate interventions to cool down our cities like planting massive tree canopies, building flood barriers, and designing new parks.

But as we looked deeper into urban planning, we discovered a heartbreaking paradox. When a city finally executes a major climate intervention in a struggling neighborhood to improve the air quality, that neighborhood suddenly becomes highly desirable. Almost overnight, property values shoot up, rents skyrocket, and the original, vulnerable residents are priced out of their own homes.

This is called "green gentrification." It means the very people who suffered through decades of pollution and lack of green space are kicked out exactly when their neighborhood finally gets fixed. We realized that cities are successfully executing climate interventions, but accidentally hurting the community in the process.

We built GreenWatch because we believe true sustainability shouldn't force us to choose between protecting the planet and protecting people. We wanted to give planners a spatial intelligence tool that uses AI to simulate the economic shockwaves of a climate intervention before it gets built, so they can finally protect the environment without leaving vulnerable families behind.

What it does

GreenWatch is an AI-powered climate equity simulation platform. Policymakers can:

Visualize 85,000+ US census tracts colored by Displacement Risk Score (DRS) and Environmental Benefit Score (EBS)
Place green infrastructure — parks, greenways, transit stops, tree planting, flood barriers, green roofs — anywhere on the map
Run what-if simulations that predict how each intervention will affect both environmental quality and displacement risk in surrounding tracts
Apply policy mitigations like rent stabilization or community land trusts to counteract displacement pressure
Get AI-optimized placement recommendations that balance environmental benefit against equity impact

How we built it

Frontend: Next.js 16 + TypeScript + MapLibre GL + Deck.gl, with PMTiles vector tiles served from Cloudflare R2 for fast, serverless map rendering
Backend: FastAPI + PostgreSQL/PostGIS for spatial queries across 85K tracts using ST_DWithin with GIST indexes
Scoring engine: Python + scikit-learn — all indicators (rent burden, PM2.5, flood risk, etc.) percentile-ranked then weighted into composite DRS and EBS scores
Data pipeline: 8-step ETL ingesting Census ACS, EPA EJScreen, CDC SVI, FEMA NRI, CDC PLACES, and Justice40 data
Simulation: Distance decay model where intervention impact diminishes from placement point to radius edge, scaled by a vulnerability multiplier
Deployment: Frontend on Vercel, scoring service containerized with Docker, map tiles on Cloudflare R2

Challenges we ran into

Scale: Processing and scoring 85,000+ census tracts with multi-source data joins was computationally intensive — we had to carefully design percentile ranking and PostGIS spatial indexing to keep queries fast
Green gentrification modeling: There's no clean dataset for displacement pressure from green infrastructure — we had to design a composite scoring system from proxy indicators (rent trends, renter %, eviction rates, market pressure changes)
PMTiles at scale: Serving a full-US vector tile dataset efficiently required migrating to cloud-native PMTiles format on Cloudflare R2 rather than traditional tile servers
Balancing equity vs. environment: Defining the right weights and multipliers for the simulation engine required significant research into existing climate justice literature

Accomplishments that we're proud of

Built a full-stack geospatial platform covering the entire United States in a hackathon timeframe
Designed a novel vulnerability multiplier that captures how the same green intervention hits already-at-risk communities harder
Two-pass simulation engine that models both intervention impact and policy mitigation effectiveness in sequence
Real PMTiles deployment on Cloudflare R2 — genuinely production-grade tile serving, not a demo mock
Integrated 6+ federal datasets into a unified scoring system grounded in climate justice research

What we learned

Climate equity isn't just about adding green space — where and how you add it matters enormously for who gets to stay
Geospatial data at national scale demands cloud-native formats (PMTiles, PostGIS) — traditional approaches don't scale
Composite scoring systems are powerful but require transparency; percentile ranking was key to making disparate data sources comparable
Policymakers need uncertainty ranges, not just point estimates — hence confidence intervals in our simulation output

What's next for GreenWatch AI

Real-time displacement tracking: Integrate live Zillow/rental market data for up-to-date market pressure signals
Scenario saving & comparison: Let users save multiple intervention plans and compare outcomes side by side
Community input layer: Allow residents to flag areas of concern directly on the map, grounding the model in lived experience
Longitudinal validation: Back-test the model against cities that implemented green projects (e.g., Atlanta BeltLine, NYC High Line) to validate predictions
Municipal partnerships: Work with city planning departments to integrate with their existing GIS and zoning workflows