The system watches your plant through BACnet sensors, learns what works with 5 ML models, experiments with setpoint changes, and keeps only what saves energy. Every cycle makes it smarter.
A continuous feedback loop that learns the most efficient operating configuration for any set of conditions
BACnet agent discovers every device on the building network. Reads telemetry from chillers, pumps, towers, AHUs, and VAV boxes every cycle.
5 ML models — baseline counterfactual, surrogate forward predictor, PCA anomaly detection, load forecast, and reinforcement learning. Ensemble disagreement provides model-free uncertainty.
Four-tier optimizer: Bayesian Thompson Sampling, coordinate descent, surrogate-first (5,000 LHS candidates), rule-based strategies.
Improvement >2%: keep. Degradation >2%: revert. Prediction gap narrows from ±12% to ±3% after 100+ experiments.
If we stop talking, the building goes back to how it was before we showed up
| Layer | Protection | Details |
|---|---|---|
| 1 | Hard Limits | CHWST 42-50°F, pumps ≥30% |
| 2 | Load Envelope | Max CHWST varies: 44°F at >80% load, 50°F at <30% |
| 3 | Return Water Watchdog | 58°F freezes exploration. 62°F emergency revert (2hr lockout). |
| 4 | Rate Limiter | Max 4°F accumulated deviation. Persisted to SQLite. |
| 5 | Dead Man's Switch | Monotonic clock. Missed cycles release all overrides. |