tell me about it, i am still working on a project where we used LLMs to assign all the UK Impact Case Studies to topics, and I then led a team doing human judgment reassignment of topics (which is so much work) but we have about 20% of misassignment by the machines for many many reasons...