Heat Threshold
Heat Threshold answers one question: should I go outside, and when?
Core trick
Computes wet-bulb heat stress from NWS and Open-Meteo data in code, then gates a go, delay, or alternate verdict behind a Gemini planner.
Description
Heat Threshold answers one question: should I go outside, and when?
You type (or speak) a location, an activity, and a time. A primary Gemini
3.5 Flash agent dispatches a graph of ten sub-agents — three uploaded as
persistent Managed Agent definitions, one weather cascade (NWS → Open-
Meteo), and six deterministic sub-agents for routing, sun-shade math,
refuge scheduling, and Street View. It returns a go / delay / alternate
verdict grounded in Stull (2011) wet-bulb math and USMC 6200.1E flag
thresholds.
Same plan, three surfaces: a 2D bento dashboard, a playable WebXR HUD
over Google Photorealistic 3D Tiles, and a Gemini Live voice mode that
calls our Managed Agents pipeline as a tool — bridging the two newest
Gemini APIs through function calling.
Every run is traced by PlatAtlas, recorded to Vercel Blob by McpTape,
and replayable via McpReplay (?replay=<runId>) — the demo safety net.
Built today, 10:30 AM – 5:00 PM PT. MIT. Environmental scheduling
tool, not medical advice.
README excerpt
# Heat Threshold **Built for Google I/O Hackathon · May 23, 2026 · 100% new work.** Heat Threshold is an environmental scheduling tool. You give it *where*, *what activity*, and *what time* — a primary Gemini 3.5 Flash agent dispatches a graph of ten sub-agents (three managed, one weather cascade, six deterministic) and returns a scheduling verdict (`go` / `delay` / `alternate`) grounded in NWS weather, USMC 6200.1E flag thresholds, and live Google Maps refuges. The same plan renders as a 2D bento dashboard, a playable WebXR spatial HUD, and a voice interface where Gemini Live calls the Managed Agents pipeline as a tool. > **Threshold is an environmental scheduling tool, not medical advice. Consult a healthcare professional for health concerns.** --- ## Gemini 3.5 Flash across every surface The same model powers every step of how this product was built and how it runs. ### 1. AI Studio (prompt iteration) The `SynthesisSubAgent` system prompt — the one that owns the medical-advice firewall, the Stull (2011) wet-bulb citation, the USMC 6200.1E flag mapping, and the strict structured-output schema — was iterated against three real input cases (SF Ferry Building, Zilker Park Austin, Hyde Park London) inside [AI Studio](https://aistudio.google.com) before being committed to code as `SUBAGENT_SPECS.synthesis.systemInstruction` in [src/lib/agents/managedAgents.ts](src/lib/agents/managedAgents.ts). The three tuned cases are the [demoFixtures](src/lib/demoFixtures.ts) presets y