HVAC Destroyer

Commercial buildings burn 40–60% of their energy bill on heating and cooling, and every building needs its own hand-tuned controls. We built one policy that runs them all. Our controller drives building HVAC every 5 minutes — AHU setpoints, dampers, reheat — balancing occupant comfort, energy and carbon in a single reward. It combines occupancy-gated rule stacks (occ / pid / bang-bang , shipped as occ_gated7) with PyTorch behavior-cloning nets (d=64 transformer over a 72-step window), scored on a deterministic simulator over full 30-day / 8640-step episodes across unseen buildings and weather. Result on NeurIPS 2026 Smart Buildings (public loss, lower is better): starter baseline 0.06439 → ours 0.01236, a 4.7× lower loss with one general policy.