Interesting direction. I’ve also been thinking about modular / subsystem-based approaches for specialized tasks in small AI systems.