Case study
An AI-powered greenhouse with public telemetry and bounded control.
A greenhouse is a useful stress test for operational AI because the world does not hold still. Weather changes, sensors fail, humidity moves, heat builds, crops have competing needs, equipment has limits, and bad actions can damage real plants.
The design principle
AI may reason and plan. AI may not directly actuate unsafe physical systems.
The system
The stack combines a physical greenhouse, ESP32 firmware, local data plane, Iris planning plane, and public proof plane. The architecture is deliberately boring around the boundary: AI proposes bounded tactics, validation checks the writes, firmware enforces physical state, and telemetry records the outcome.
The honest claim
We are not claiming the AI magically caused every improvement. We are claiming the system makes planner availability, stress, compliance, resource use, and score visible enough to audit.
| Metric | Planner offline | Iris online |
|---|---|---|
| Average Iris plans/day | 0.0 | 3.0 |
| Both-axis compliance | 20.1% | 54.7% |
| Temperature compliance | 45.3% | 70.5% |
| VPD compliance | 30.1% | 74.4% |
| Stress-axis hours/day | 29.8h | 12.9h |
| Water/day | 427.5 gal | 222.7 gal |
| Estimated electric energy/day | 2.6 kWh | 1.2 kWh |
| Cost/day | $6.05 | $5.12 |
| Planner score | 28.0 | 56.9 |
Caveat: this is operational comparison evidence, not a controlled A/B test. The Iris-online window was cooler, more humid, and lower solar, so the correct claim is auditability and operational evidence, not isolated causal proof.
Business lesson