I fly commercially. I've sat in the flight deck through ground stops, ATC reroutes, and cascading network delays across hundreds of sectors. I know exactly what causes a late departure — and more importantly, I know when it's avoidable.
DelayGuard started as a frustration I couldn't fix from the cockpit. Passengers would board already stressed about connections, often on flights I could have told them — hours earlier — had a high chance of delay. No consumer tool existed to give them that information before they booked. FlightAware tracks delays after the fact. Flighty works day-of on iOS. Nothing gave a pre-booking probability for any flight, on any date, on the web.
So I built one. I combined real BTS on-time performance data with the operational patterns I understand from flying — how hub congestion compounds, how weather at one airport cascades to another three time zones away, how a 6am departure behaves completely differently to a 6pm one on the same route. That expert layer is baked into the model's pre-departure signals.
The result is XGBoost v9: trained on nearly 4 million real US domestic flights, using only what's known before departure, and tested on later flights it never saw — built by someone who's lived the delays, not just modelled them.