mullama is a research project exploring what a unified local LLM serving layer looks like when built from first principles. Rather than wrapping existing tools, it directly interfaces with llama.cpp to study model lifecycle management, inference scheduling, and API compatibility.