SkyRipple simulates exactly that - live, against a full real day of US domestic air traffic - then deploys a five-role AI operations team to recover it and shows you the dollars saved.
582,304
real BTS flight legs
Dec 2025 · 343 airports · 14 carriers
5.19x
cascade propagation
$181,376 direct → $942,081 network-wide
2.41M
synthetic passenger itineraries
190,711 crew pairings, 6,756 gates
16-20s
full national cascade
steady-state live resolution time
Four systems working together: a world model built from real data, an engine that propagates disruption through it, a ledger that prices every consequence, and an agentic team that recovers from it.
A four-channel simulation - aircraft rotation, crew pairing, passenger itineraries, and disruptions - propagates one localized event (a runway closure, a grounded tail) into its full network-wide consequence across a real day of US domestic traffic.
Every dollar is attributed down to the carrier and airport across five cost categories, validated against real BTS on-time data (reactionary delay correlates r ≈ +0.6 with the ground-truth late_aircraft_delay_min).
Five LLM-powered roles - Aircraft, Crew, Passenger, and Gate Controllers, arbitrated by a Duty Manager - propose recovery plans that are re-simulated and cost-checked before acceptance. The arbitrated plan can never cost more than doing nothing.
A FlightRadar-style live map, a scrubbable timeline, and a real-time cost ledger let you provoke a disruption and watch it ripple through the network minute by minute - not read about it after the fact.
The full engine keeps a live Python discrete-event simulator, a real SQLite world model, and Gemini-driven natural-language parsing running behind a FastAPI backend - a five-minute wait for an enterprise-grade calculation. This Lite build is that same engine's output, pre-computed and packaged as a zero-latency static client, so the analytical depth is instant to explore.
Why I built it this way