Scaling cherry tomato production without scaling labor

Case study

How AI-Driven Light Control Raised Lettuce Yields 44%

Lettuce

Deep Flow System

4,960 m² Facility

Gimje, Korea

Key Results

+44% improvement in yield
+40g increase in weight per head
24 rows, 2,300 plants/row

About the facility

Where Croft's earliest work began

The earliest version of Croft's approach was tested and refined in a 4,960 m² facility allocated through a Gimje government division, not selected for growing conditions but assigned as part of the program's structure. Croft's CEO, Hee Kyung, chose lettuce because she knew some of the most robust technology in agriculture was already being used to grow it. Still, the facility fell short of some of the most essential requirements for growing it well.

JE Farms: cherry tomato production at scale

The challenge

Couldn’t grow in the two seasons with higher prices

Lettuce root systems are highly sensitive to water temperature. Below roughly 15°C, roots stop absorbing nutrients from the solution, and growth stalls regardless of what's happening abve the surface.

The facility had no chiller and no boiler, so there was nothing keeping the root-zone temperature inside the ideal ranges.

This meant only spring and fall were commercially viable for growing. Summer ran too hot, winter too cold. Even worse, these two seasons are when lettuce prices are highest, and they were the two windows that lettuce couldn’t grow in this facility.

Our solution

Light control that follows the plant's lead

The team refined a Deep Flow System (DFS), adjusting the crop's environment to the plant rather than a fixed schedule. Light was the primary lever: too much direct sun in the afternoon pushes lettuce to burn stored sugars, protecting itself from stress rather than growing. Shade screens were automatically adjusted ahead of those conditions, matching the crop's growth stage and the greenhouse's temperature.

Irrigation followed automatically from there; driven by the cumulative light the crop had received rather than a fixed schedule. The system trained continuously throughout the trial rather than running on a single fixed setup, building on an expert-defined baseline strategy with specific boundaries designed to keep the AI from overcorrecting. Ultimately, the approach directly resulted in a 44% increase in growth yield over the trial period and was applied as an ongoing solution.

The results

A 44% gain despite facility limitations

Over the 27 January–3 March 2023 trial (measured against a baseline cycle grown at the same time of year, at the same facility), fresh weight per head rose from ~90g to ~130g: a 44% increase, or roughly 4.97 to 7.18 tonnes per cycle across the full greenhouse. It's a controlled before/after at one facility, not a peer-team benchmark, which is exactly what makes the number solid.

+44%

Improvement in yield volume

+40g

Increase in average fresh weight per head

+55,000

24 rows x 2,300 plants each, full commercial-scale

The impact

Proof our business model could generalize

The real result here wasn't the yield number; it was proving our approach would generalize. A control approach built on data-driven optimization with a safety net of expert grower knowledge held up in a facility never built for ideal conditions, on a crop that wasn't hand-picked for it. That's the test a platform needs to pass, not a best-case demonstration. The same principle later carried into Croft's trellised tomato system, proof that this wasn't a one-off, but a philosophy we could build on — one we are still refining today.

2.5+

hrs/week saved per grower