About the project

Inspiration

Extreme heat events are becoming more frequent, but response planning is still often reactive and city-wide instead of neighborhood-specific.
We wanted to build a tool that helps emergency teams answer one critical question quickly:
Where is the risk highest right now, and what should we do in the next 72 hours?

What we built

We built HeatShield AI, a real-time municipal heat response platform that:

• Pulls live weather and AQI signals
• Builds neighborhood-level risk views from open map data
• Simulates intervention impact (cooling centers, outreach, mobile clinics)
• Generates a structured 72-hour response playbook with Gemini

If model generation fails, the app automatically returns a deterministic fallback plan so operations can continue.

How we built it

• Frontend: TanStack Start + React + Tailwind CSS
• Backend: TanStack Server Functions + Nitro Node server
• Data: Open-Meteo (weather/AQI), OpenStreetMap Overpass (neighborhood/civic features)
• AI: Gemini API with multi-model failover
• Reliability: in-memory caching, source tracking, and graceful fallbacks

Risk scoring combines heat pressure, vulnerability, and intervention effects.
At a high level:

$$ R_{adj} = \max\left(0,\; R_{base} - \Delta_{intervention}\right) $$

where:

$$ R_{base} = \alpha \cdot H + \beta \cdot V + \gamma \cdot U $$

• (H): heat pressure (temperature, humidity, AQI, grid stress)
• (V): social/health vulnerability (age, chronic illness, poverty, tree cover)
• (U): urban heat island component

What we learned

• Live public data is messy; robust fallbacks matter more than perfect inputs.
• Model reliability is an engineering problem, not just a prompt problem.
• Clear UX for uncertainty (data source labels, fallback notes) builds trust.
• Fast iteration is easier when frontend and backend share one typed stack.

Challenges we faced

• Model quota/rate limits: solved with model failover and transparent error reporting.
• Geospatial inconsistency: many places have sparse map tags; solved with derived features + synthetic fallback neighborhoods.
• Operational usability: raw LLM output was hard to consume until we added markdown + math rendering and structured sections.

What’s next

• Add historical trend layers and forecast confidence intervals
• Add role-based operator workflows (health, utilities, emergency management)
• Integrate official city datasets for shelter occupancy and ambulance pressure
• Export machine-readable action plans for incident command systems

Built With

• gemini
• openmeteo
• openstreetmap
• react
• tailwind
• tanstack