.jpg)
The farming population in South Korea is aging fast. The gap between retiring growers and new ones entering the field has real consequences for food security, and I decided to find out what that gap looked like from the inside.
In 2020, I enrolled in one of the Korean government's two-year agricultural training programmes: a structured education in greenhouse cultivation alongside other young farmers. We studied growing in a greenhouse through books and theory for two months, then spent six months working inside a commercial greenhouse alongside professionals, and finally spent a year managing a 1,500 m² government-allocated facility with a team of four or five students. We got to choose which crop we were interested in growing; I chose lettuce.
But as the growing portion started, a huge problem became clear. The facility hadn't been built by professionals who understood what growing lettuce actually requires. Lettuce root systems are highly sensitive to water temperature. But there was no chiller, no boiler, no system to control root zone temperature. When the water drops below 15 degrees, the roots can't absorb nutrients from the solution. This means it’s impossible to grow profitably in winter; it’s too cold. Summer is also too hot. The two seasons when the price of lettuce was the highest were the two seasons we couldn’t grow lettuce.
This wasn't unique to my cohort. Around 200 students complete this programme every year. Of those who go on to build their own greenhouses, most fail in their first growing cycles. The knowledge required to manage a controlled environment successfully takes three to five years to develop and can't be transferred through a two-year programme. Get one setting wrong at week two, and you might not find out until week six, when the crop is already lost.
But I realized that if that knowledge could be encoded and more of these decisions could be automated, the expertise problem stops being a people problem. It becomes an engineering challenge. The question was whether anyone could build it.
Woo Ram and I had met at university, in the same beginner swimming class. We'd stayed in loose contact in the years after, but by the time I called him I was only partially aware of how far his work had taken him.
His PhD focused on recording and classifying olfactory nerve signals in mice using machine learning. The method: implant an electrode, record biological signals as the animal encountered a smell, train an algorithm to identify which compound triggered which response. A living system, constantly producing data. The question was whether you could learn to read it accurately enough to understand what it meant. After his doctorate, he moved into medical image analysis, using light spectroscopy and deep learning to assess the cleanliness of medical devices. Different application, same underlying logic. A biological system producing signals, an algorithm learning what those signals mean.
When I described what I'd been watching in the greenhouse, Woo Ram recognised the structure immediately. A plant expresses information continuously. Every shift in leaf colour, stem angle, or growth rate is a signal. What it needs, what it's missing, what's failing. Most growers were reading some of it through instinct that couldn't be transferred or scaled. A system trained carefully enough could read all of it, consistently, across every hour of every growth cycle, and act on what it found without waiting for a human to notice.
If that could be built, we’d have a real solution to the knowledge gap I recognized.
We needed a testing ground where this thesis would either hold or break against real conditions.
The Autonomous Greenhouse Challenge at Wageningen University was it. The competition draws teams from across the world to grow crops fully autonomously in real greenhouse compartments. Algorithms in control, no human intervention once the growing period begins. If our approach held up there, it held up anywhere.
.jpg)
We entered the third edition as Team CVA, Crop Vision and Automation. The crop was lettuce. Roughly eighteen months to design and build something that could perform in real conditions.
Plants grow on their own schedule. That was the thing we couldn't engineer around. We were using a moving trough system, which gave us a continuous range of growth stages to learn from, but fine-tuning still required following individual plants from seed to harvest. The crop confirmed whether we were moving in the right direction. Not the model, not the data. What we slowly understood was that training these models is less like building infrastructure and more like growing the plants themselves. You set the conditions, make your best decisions, and wait.
The competition opened with a 24-hour hackathon. Seventeen teams, more than 140 participants across 18 nationalities, each judged on how profitably their AI could grow lettuce inside a sophisticated greenhouse simulator built to behave the way real crops do: unpredictably.
We finished first, with more than double the score of the nearest competitor. It was the first real signal from leading researchers in the field that our methodology was sound. We had spent over a year building something that felt right but hadn't been validated outside our own work. That result told us we weren't wrong about the approach.
Then came the growing round. From February 2022, the five finalist teams each had a real greenhouse compartment at Wageningen's facility in Bleiswijk. Real lettuce. Real sensors. Algorithms running the climate control systems are responsible for every decision once the doors closed. Every team pretty much did the same thing: took the algorithm that had won in simulation and applied it directly to a living crop.
Every team finished below their simulation results. We dropped from first place with a long lead to finishing third overall.
We went back again. In the fourth edition of the challenge, competing as Team AgroFusion, we tried again to refine our algorithm. This time, we were growing cherry tomatoes, a crop we hadn’t grown commercially before. This time, we finished second in the growing phase, but the results were still not limited by needing to just program the automated system without feedback from the plants.
What we realized across these two challenges was that a simulation cannot account for what a living plant does. Seed batch variation, micro-environmental differences, and biological responses to conditions are too diverse for any model to fully capture. An algorithm built on prediction will always hit the limit of what we are able to predict in a simulation. We needed to understand how the plant is behaving.
Bleiswijk produced the insight that became the core of everything we've built since. In agriculture, using AI is not about predictions; it's about responsiveness. You have to be reactive to what the plant is showing you, not just model what you expect it to do.
Coming back to Wageningen University last month, I found myself walking through some of the same spaces.
The competition didn't start Croft. Seeing how hard it was for young growers had already convinced me that the work we were doing was necessary. But competing confirmed what we needed to hear from the market: that the problem was real, the approach was logical, and the researchers who had spent careers on the hardest problems in agricultural technology thought what we were doing was worth taking seriously.
The system Croft operates today is built on everything we learned across those years and the two growing challenges. But the principle that drove us into the competition is the same one that drives the system now: making growing food as profitable and scalable for farmers as possible.
%20Croft%20Logo_(Long).png)