A solo developer used Seedance, FFmpeg, and real screen recordings to give Amana two AI spokespersons without letting the model fake its UI.