STRATA

Inspiration

Every sustainability score tells you where something stands today — none of them tell you which direction it's heading, or how much to trust the number. A neighborhood can score green on every index while rent doubles and long-term residents are displaced. A company can publish a net-zero pledge while its Scope 3 emissions keep rising. We wanted to build something that grades the trajectory, not just the present state — and is honest when it doesn't know.

What it does

We learned that the most valuable signal in sustainability analysis isn't a score — it's disagreement. When agents diverge sharply, that divergence is itself actionable information no rating agency produces. We also learned how to architect multi-agent streaming pipelines with asyncio.gather() and push live token-by-token updates to the browser over SSE.

How we built it

STRATA is a dual-mode AI platform powered by Gemini 2.5 Flash. Five specialized agents fire in parallel — each reasoning through a different lens (climate, equity, operations, compliance, and a Devil's Advocate). Results stream live to a React frontend over SSE, updating a radar chart token by token as each agent finishes. All data sources are free: Sentinel-2 satellite imagery, Census ACS, OpenAQ, EPA, SEC EDGAR, and more.

Challenges we ran into

The hardest challenge was the green gentrification detection — cross-referencing every positive sustainability signal against Census rent trajectory to flag when improvements displace the residents they're supposed to help. Another challenge was designing the Devil's Advocate agent to always challenge the most confident claim with a specific cited counter-source, rather than just adding generic skepticism.

Accomplishments that we're proud of

-Built a fully working multi-agent debate architecture where 5 Gemini agents fire in parallel and challenge each other's conclusions in real time

Implemented green gentrification detection — the only sustainability tool we know of that flags when environmental improvements displace the residents they're meant to help
Designed the CONTESTED verdict, which surfaces agent disagreement as a first-class output rather than hiding it behind an averaged score
Achieved real-time token-by-token streaming to a live radar chart using SSE — the UI updates as each agent thinks
Built entirely on free APIs with no paid data subscriptions beyond the Gemini API key

What we learned

Multi-agent disagreement is more informative than consensus — CONTESTED is often the most honest and useful verdict
Streaming AI responses to live visualizations requires careful state management between backend SSE and frontend chart updates
Green gentrification is systemic and measurable — rent burden data from Census ACS tells a story that satellite NDVI scores completely miss