Turning a heritage university telescope into a taskable sensor, and marginal optical tracks into orbit solutions with honest uncertainty.
Telling a real manoeuvre apart from mismodelled dynamics or a bad track, across the GEO belt, using labelled optical data rather than assumptions.
Reconnecting fragmentation debris to the event that produced it, by widening the orbital fingerprint a classifier can see.
Cutting the dominant error source in low-Earth-orbit prediction by learning reduced-order density models from real ephemerides.