I built a dual RTX 3090 rig for local AI in 2025 (and lessons learned) (www.llamabuilds.ai)