A lot of people focus on picking better models but ignore memory design. In production, small things like bounded caches and cleanup rules make a huge difference.