0Nyquestbycompetitive_fuchsia_hermioneOptimize LLM token usage with effective semantic compression.GitHubView pitch