Understanding where and how marine predators forage is fundamental to effective conservation, yet direct behavioural observations are typically restricted to short periods of animal-borne video. We developed a hierarchical machine-learning framework to extend behavioural inference beyond observed footage in the endangered Australian sea lion (Neophoca cinerea). Animal-borne video was integrated with head-mounted accelerometer and magnetometer data, depth, and movement context from eight adult females instrumented at two South Australian colonies. Using 2 s overlapping windows, a first-stage random forest model distinguished foraging from other behaviour with strong transferability across individuals (leave-one-individual-out cross-validation accuracy = 76.4%). Building on this foundation, secondary models were applied within predicted foraging periods to resolve finer behavioural states, foraging outcome, and prey type. This staged approach aligns the modelling framework with the biological structure of the behaviour itself: first identifying when foraging occurs, then asking what type of foraging event is taking place. The resulting workflow transforms short-duration video records into broader behavioural predictions across complete foraging trips, creating new opportunities to map activity in space and time from multi-sensor biologging data. These preliminary findings highlight the promise of hierarchical AI for intelligent marine monitoring and demonstrate a scalable pathway for improving conservation-relevant knowledge of habitat use, foraging ecology, and management needs in marine megafauna.