A vision-language model reviews drone footage frame by frame, boxes the damage it sees, and drafts a pre-inspection summary for the adjuster.