SuperScade

Inspiration

In a disaster, the first one to two hours decide who lives. Yet the data a responder needs in that window sits in a dozen siloed systems that don't talk to each other:

• Hazard feeds: USGS ShakeAlert and earthquake feeds, NWS/NOAA warnings, CAL FIRE incidents, USGS stream gauges
• Dispatch and coordination: 911 CAD systems, WebEOC incident boards, Cal OES, FEMA
• Infrastructure: Caltrans/511 road and bridge closures, PG&E gas and power outages, water utilities
• Medical and shelter: hospital bed status (EMResource), Red Cross shelters
• City data: DataSF building, permit and population data, 311

An incident commander pieces this together by phone and radio while the clock runs. Palantir did this fusion for military data. Nobody has done it, agent-native, for the golden hour of a civilian disaster.

What it does

SUPERCADE turns one disaster into a live, verified response plan in under a minute.

• Trigger a scenario, for example a 1906-style M7.8 on the San Andreas. Cascade rules set off the knock-on events: liquefaction in the Marina and SoMa fill, Golden Gate and Bay Bridge damage, gas-line fires, aftershocks and Presidio landslides.
• A Chief agent reads the situation and spawns one agent per district (11 San Francisco neighborhoods). Each district agent spawns its own helpers for routes, fire crews, shelters, airlift and supplies.
• An independent Verifier rejects any plan that leaves people cut off or a fire without crews, and the district re-plans live.
• A risk monitor projects the next hour and escalates districts: monitoring, alert, evacuate, escalate. Escalated districts phone their emergency contact.
• After every run a Reflector rewrites the team's playbook. The change is kept only if it scores better on held-out scenarios, so the system gets better with every disaster.
• Everything plays out on a 3D map of San Francisco and the Bay Area with real highways, terrain, spreading fires and shaking.

How we built it

• Recursive agent tree on Crusoe Managed Inference: NVIDIA Nemotron 3 Super 120B (Chief, Verifier, Reflector), Nemotron 3.5 Lightning 30B (district agents) and Nemotron 3 Nano (helpers).
• You.com Search pulls real-world facts (hospitals, bridges, past events) into a shared memory every agent reads.
• The deterministic simulation engine handles cascades, routing, scoring and held-out evaluation. The verifier combines simulation checks with LLM review, so "done" means the plan actually works in the sim.
• The 3D client uses TypeScript, React and Vite, with a live event stream from a Node server.
• A live run on San Francisco: 11 district agents dispatched at tick 1, an M6.6 aftershock at tick 6, and fires spreading through Twin Peaks, Downtown and the Western Addition.

Built today

The San Francisco region: 11 neighborhood districts with real bay-fill, hill and coast terrain, the San Andreas M7.8 cascade with a Diablo-wind fire, real highway and arterial centerlines, a low-detail surrounding Bay Area (Marin, East Bay, Peninsula, the San Andreas trace), a terrain legend and a cityscape scene. These sit on top of the recursive agent engine I built at the Crusoe Recursive Agents hackathon on Oct 17.

Challenges

Keeping 40+ live model calls fast enough to plan inside the golden hour, making the verifier reject bad plans without blocking every plan, and modeling San Francisco's real liquefaction zones and road network at neighborhood scale.

What's next

• Plug in the live feeds above (USGS, NWS, CAL FIRE, 511, PG&E, hospital status) in place of simulated inputs.
• Real telephony for the district calls.
• Pilot with a county Office of Emergency Management as the agent layer on top of tools like WebEOC.

Built With

• crusoe
• docker
• node.js
• nvidia-nemotron
• react
• three.js
• typescript
• vite
• you.com