I was going to suggest chess position recognition, AFAIK it's a completely unsolved computer vision task (once a position is recognized, I think analysis is well solved by, say, a stockfish tool for the LLM, but there is interesting work going on with language models themselves understanding chess)