How LLM helped crafting a tailor-made compression algorithm for our temperature/humidity time series data, compressing it by ~x53.6.