Observe
Overhead cameras read the canopy every ninety seconds — leaf angle, color drift, the first millimeter of wilt. Nothing taped to a stem. The plant itself is the instrument.
phase i — the lookingGrowth intelligence for glasshouses. An AI that tends living crops the way a gardener does — by watching, all night, leaf by leaf.
A night shift made of cameras and climate loops. It reads the canopy, steers the air in millimeters, and knows the harvest window before the flower does.
Walk the naveNot a dashboard bolted onto a greenhouse — a closed loop that looks first, decides second, and touches the crop last, as lightly as it can.
Overhead cameras read the canopy every ninety seconds — leaf angle, color drift, the first millimeter of wilt. Nothing taped to a stem. The plant itself is the instrument.
phase i — the lookingA growth model trained on twelve simulated seasons weighs light, vapor and root moisture, then picks the smallest intervention that keeps the curve on track.
phase ii — the weighingVents, shades, drip lines and fans move in millimeters, not lurches. Every action is logged, reversible, and small enough to undo before morning.
phase iii — the touchingEach module does a single job well, and they argue with each other before anything in the glasshouse moves.
Reads turgor, chlorosis and tip growth from overhead scans. In our simulations it flags a wilting row eleven hours before a walking inspection would.
Weaves light, humidity and airflow into one loop, holding vapor-pressure deficit inside a hair's width of target through a desert night.
Models the root zone as a moisture-and-nutrient graph, and predicts the dry pockets a drip line will miss before it misses them.
Samples the air for spore load and quarantines a bay with airflow alone — no walls, just a wall of wind.
Calls the harvest window with its error bars printed as large as the number. When the bars widen, it says so first — before you ask.
Every forecast ships with its uncertainty, and the uncertainty is typeset as large as the number. A confident wrong answer is the only failure mode we refuse to tolerate.
When the model and the leaf disagree, the leaf wins and the model retrains that night. No sensor reading outranks the crop it was pointed at.
Any grower can pause the entire system from the aisle with one hand, and it stays paused until a human — not a scheduler — turns it back on.
Every vent that opens traces back to the readings that opened it. If we can't show the chain, the vent doesn't move.
The whole system speaks a small protocol. You ask a nave what it sees; it answers with a growth index, a stage, and the smallest useful action — never a bigger one.
The same index drives the meter in the corner of this page. Scroll, and you are walking a season.
// ask the east nave what it sees POST /v1/naves/east/observe { "scan": "0412-night", "resolve": ["turgor", "chlorosis", "tip_growth"] } // it answers with the index — and its doubt 200 { "growth_index": 0.62 ± 0.04, "stage": "CANOPY", "advice": [{ "actuator": "vent_3", "action": "open_4mm", "because": "vpd drifting high since 02:10" }] }
We take on one glasshouse a season, and we stay through the harvest. Tell us what you grow and what keeps you up at night.
Write to the greenhouse hello@verdantia.ai