I built a greenhouse where an AI agent (OpenClaw) planner optimizes climate changes and ESP32 firmware controls the relays
A 367 sq ft greenhouse in Colorado where AI plans but never touches the hardware...
The Verdify project demonstrates a practical safety boundary for AI in physical systems. Built in a real 367 sq ft greenhouse in Longmont, Colorado, the system collects telemetry including temperature, humidity/VPD, equipment state, resource usage, weather context, and scorecards. An AI planner (OpenClaw) analyzes recent conditions, plant target bands, equipment limits, and forecasts to propose bounded 'tunables' for firmware enforcement. Every proposal passes through a dispatcher that validates schema, checks bounds, clamps invalid values, and rejects anything outside the safety envelope.
The ESP32 firmware maintains full control over relay loops for fans, misters/foggers, and heat. The key insight: AI suggests optimal adjustments to save water, electricity, and gas, but never directly flips a relay. This architecture ensures plant climate stability while letting the AI optimize tradeoffs. The project is open-source on GitHub with detailed safety documentation and video overview—a model for safe AI-controlled environments.
- OpenClaw AI planner proposes climate adjustments (tunables) but cannot directly control relays
- Dispatcher validates and clamps all proposals before ESP32 firmware executes relay actions
- Real 367 sq ft greenhouse in Colorado with open-source safety architecture on GitHub
Why It Matters
Shows a practical, safe pattern for using AI in physical systems without bypassing safety-critical controls.