How we saved 80 % off LLM inference costs by pruning “junk tokens”. And why this pattern works almost anywhere.