From hallucinations to hardware: Lessons from a real-world computer vision project gone sideways