“Philip Kelly's book, "Inference Engineering," posits that AI inference is a full-stack problem combining CUDA, on-GPU optimization, and distributed systems.”