Verdify.ai — AI Proposes. Firmware Controls. Telemetry Verifies.

Added: 2026-05-17

[00:00:00]We built Verify to answer a practical question.

[00:00:03]Can a real greenhouse, managed by an AI agent, stay closer to what the plants need while using water, electricity, and gas more intelligently? Simply put, how [snorts] can we maximize efficiency and reduce cost using AI in our greenhouse?

[00:00:20]I'm Jason Valerie. And I'm James Valerie. We built Verify together in Longmont, Colorado. It is a 367 square foot greenhouse with climate sensors, fans, misters, a fogger, heaters, and an AI planning loop powered by an Open Claw agent and enforced by an ESP32 microcontroller.

[00:00:42]The goal is to optimize a living environment under real constraints. The plants care about the temperature, humidity, VPD, light, air flow, wet and dry cycles, and stress hours.

[00:00:55]Resources matter, too.

[00:00:57]Misting uses water, fans use electricity, and heat uses gas.

[00:01:02]Verify is designed to balance both sides, the plants' climate quality and resource cost.

[00:01:09]We did this because greenhouses are constantly changing. Time in the sun, outdoor climates, and specific environmental characteristics affect how a greenhouse behaves. Because of this, the question what's best for optimization isn't what's best right now, but what's best in aggregate.

[00:01:24]That's where the AI comes in. It looks [snorts] at recent sensor data, plant requirements, target bands, known equipment limits, and weather forecasts.

[00:01:32]Then, it can send targeted adjustments to the controller. We call those adjustments tunables. A tunable is not a relay command. Rather, it is an adjustable parameter inside the controller. For example, temperature targets, VPD bands, mister schedules, fan thresholds, and hysteresis are all tunables. It can't say, "Turn the fogger on forever." The controller has hard rules in place to prevent this. Every proposal goes through the dispatcher that validates the format and checks to see that it remains in the safety boundaries in place.

[00:02:02]Then, the SP32 validates the greenhouse state, checks the approved tunables, and decides in real time what to do the equipment should do.

[00:02:11]That's dynamic greenhouse control.

[00:02:13]And the Colorado climate is just as dynamic and brutal.

[00:02:17]We experience extreme versions of all four seasons, sometimes on the same day.

[00:02:23]The sun load changes, the outdoor air changes, humidity changes, plants change. A fixed rule could work in the morning and waste water or power by afternoon.

[00:02:34]Verdify treats each plan as a hypothesis. The AI planner runs at sunrise, solar maximum, and sunset each day.

[00:02:43]The planner is also dynamically triggered if the system identifies a substantial deviation from the established weather forecast, enabling the greenhouse to respond to unexpected changes in the outdoor weather.

[00:02:57]The AI planner suggests a climate tactic. The controller runs safely and the telemetry comes back.

[00:03:04]The scorecard tells us whether we improved the environment, wasted resources, or made the wrong set of tradeoffs. Over time, the AI learns from its past experiments. And this is why we made the project public. We published the planner outputs, the telemetry, the score cards, the costs, the failures, the known limits, and static data samples for analysis.

[00:03:28]That's Verdify. The data coming comes in, the AI proposes, firmware controls the equipment, and then the cycle repeats. The site is live at verdify.ai.

[00:03:37]We built this together and we would love serious critique from people working on embedded systems, controls, greenhouse operations, and local AI.