S.01input
Address Input
User enters a real location or planning goal.
[003]·jun_2026v1.0
UrbanPilot
Decision-support prototype. Results depend on public-data coverage, scoring assumptions, geographic availability, and source freshness.
Multi-agent AI system that runs parallel scenario analyses — climate, housing, accessibility, urban design — for any real address.
AgentsClimate / Housing / Accessibility / Urban Design
SourcesACS / FEMA / Open-Meteo / 511 Transit / NLCD / Maps
MethodParallel scenario analysis
OutputRisk + recommendation dashboard
What I Contributed
- Built the React (Create React App + Tailwind + Leaflet) frontend and the Node/Express backend.
- Built the Claude multi-agent coordinator orchestrating the climate, accessibility, housing, urban-design, and vision agents in parallel.
- Integrated Midjourney (via MCP) for 2040/2075 scenario visualization, grounded in a real Street View or satellite photo of the site.
- Authored the prompt-compression benchmarking: a hand-written compact-encoding layer plus a the-token-company integration, measured against real Anthropic API calls.
- Imported and substantially rebuilt an earlier hackathon prototype (GreenPlanner, HackMIT) into UrbanPilot.
- Teammate kdai05 built the verified-data grounding integrations for each specialist agent (Census ACS, FEMA, Open-Meteo, 511 GTFS, NLCD) and grounded the Ask-AI assistant and vision scenarios.
How It Works
4 stagesS.02etl
Data Pipeline
System pulls and normalizes public datasets across climate, transit, housing, and hazard layers.
S.03agents
Multi-Agent Analysis
Specialized agents evaluate tradeoffs independently, then combine findings into scenario scores.
S.04output
Planning Output
Dashboard surfaces risks, recommendations, and long-term scenario comparisons.
Why It Matters
Planning decisions trade off development, climate exposure, mobility, and affordability — usually across tools that don't talk to each other.
UrbanPilot puts those tradeoffs in one view. Planners, students, and community teams enter a real address and get parallel agent analyses grounded in verified public data — enough to compare scenarios and defend a recommendation in minutes, not weeks.
Evaluation & Evidence
- Prompt-compression benchmark against real Anthropic API calls (no estimates): the hand-written compact encoding cut input tokens by 27.8% (Housing agent, 1,155→834), 35.4% (Climate, 1,507→973), and 33.7% (Accessibility, 1,247→827), with every compressed response still parsing as valid JSON and citing the exact verified source figures.
- A separate the-token-company compression layer reduced a structured JSON-schema prompt by 3.9% and a long prose-context prompt by 20.6%, layered on top of the encoding above.
- Verified graceful degradation: without a Census or 511 API key, the corresponding agent still runs without that verified grounding rather than failing; without an Anthropic key, that section shows as unavailable rather than a fabricated result. There is no bundled mock data anywhere in the app.
Constraints & Limitations
- No automated test suite (the repo's own README notes `npm test` reports "No tests found").
- 511 GTFS transit grounding is Bay Area-specific; coverage elsewhere depends on the other four sources alone.
- Midjourney can't fetch localhost/LAN reference photos in local dev — the "use the real photo as reference" feature only works once deployed with a public URL.
- Google Maps imagery is only ever displayed, never used as Midjourney training/input data, except the explicit reference-image feature above.
What I'd Improve Next
- Add an automated test suite.
- Extend verified transit grounding beyond the Bay Area.
- Verify the Midjourney reference-photo flow end-to-end once deployed.