Skip to content

Storm simulation agi

An El Niño storm is coming for San Francisco.

Code Video Cerebral Valley

Core trick

Simulates flood and wind damage across a real city slice, then Argon proposes response plans that are re-scored by the same simulator.

Tags

science

Description

An El Niño storm is coming for San Francisco. StormSim SF simulates it on real city data: flooding streets, overwhelmed drains, wind toppling trees, and the dollar damage to every building. Gemini 4 Argon gets the whole city and a $50M budget, and plans the response like a city emergency team: floodwalls, pumps, sandbags, shelters and evacuation. It runs the storm, watches what happened, and improves its plan. Result: damage fell from $258M to $18M with a $43M plan.
TurpoFIN/argon-hackathon★ 00 forksPythonCreated Oct 25, 2026 · last push Oct 26, 2026

README excerpt

# StormSim SF: Gemini 4 Argon defends San Francisco from an El Niño storm

An El Niño atmospheric river is heading for San Francisco. StormSim SF simulates it hitting a real 3.5 × 4.3 km slice of the city (downtown towers, SoMa and the Mission Creek basin around 17th & Folsom): rain, runoff, overwhelmed storm drains, a surging bay, wind through the tower canyons, street trees snapping and uprooting, cars floating, and 250,000 people. It prices the damage in dollars.

Gemini 4 Argon then gets the whole city in its 1M-token context and a budget, and designs the city's response the way a planning cell would: floodwalls and gates, pumps, cisterns, a relief tunnel, permeable streets, drain clearing, sewer drawdown, shelters, evacuation of vulnerable residents, food and water logistics, and crews. It writes the plan as code against a city API, builds it in 3D, runs the storm, watches the replay video, critiques its own plan and fixes it. Every model gets the same storm, data, tools and budget. The planner approves the final Emergency Operations Plan.

**Problem statement:** Two, More Dimensions (generate a 3D world, render it, watch it, critique it, fix it in a loop).

## What's real

- **Data:**
  - USGS 3DEP lidar terrain.
  - DataSF building footprints with lidar heights, 23,564 street trees, 4,982 storm inlets, assessor parcels, census population and flood-health vulnerability.
  - The SFPUC 100-year storm flood map.
  - 311 reports from the 31.12.2022 storm.
  - NOAA tides, At

Team