Recovery agents for airline irregular operations, reasoning over a FalkorDB graph of real US DOT flights. They propose; you approve; the graph checks every change.
In the week before Christmas 2022 Southwest cancelled nearly 17,000 flights. Its crew system lost track of where crews and aircraft were and how they connected. This replay loads the airline's real schedule for 21–23 December from US DOT data, and lets three agents work the storm from the graph.
The graph paths behind the agents' answers appear here.
Nothing changes the operation until you approve it.
Current FAA airport status, METAR weather and the next 12 hours of TAF forecasts, written into the airline's graph and matched against its route network.
US DOT records how many minutes of each flight's delay came from its aircraft arriving late. We seed the graph with only each flight's own weather, air-traffic and carrier delay, let it propagate along aircraft rotations, and compare.
Each scenario is run as the agents would at its clock. A cancel is borne out if the airline also cancelled the flight or flew it 3h+ late. Projected 3h+ delays (the 14 CFR 260.6 refund line) are checked against real arrivals, and compared with the base rate among every flight downstream of the known delays.
The read-only tool catalogue is also an MCP server: the same human-written Cypher, id resolution and paths, for
Southwest (WN) and Alaska (AS), plus the decision log, the cited rules and past lessons as
resources, and a briefing prompt. Add it to Claude Code:
claude mcp add --transport http irops-brain /mcp/