Compression layer that strips useless tokens from LLM prompts in milliseconds, reducing costs and improving output quality.