When David Ahmed was a student at New York University, he often watched the growers coming and going at Manhattan farmers markets. Sometimes he would leave with an empty truck. Other times, the vehicles were still full of produce.
Believing that there is a better way to produce food and, more importantly, a better way to predict how much farmers actually need to produce, Ahmed launched Felix in 2019.・Established Hexafarm with Kirschstein, Kim Hui-jo, and Abraham Dul.
Today, the company uses sensor technology and AI to track the entire agricultural environment, providing yield forecasting, climate monitoring, fruit counting, and more for greenhouse and indoor farm growers. Its long-term goal is to build machine learning models that can surpass humans at optimizing production in any growing environment.
“The system is very dynamic,” says Ahmed. agfunder news. “Just from the data collected, we can tell growers how much they have harvested. [will] Harvest a few weeks in advance. Since we know this, the next step is to provide producers with viable strategies to get the most out of the resources they put into it. ”
Hexafarms, based in Berlin, Germany, just raised €1.3 million ($1.4 million) in pre-seed funding led by Speedinvest, with participation from Mudcake and Techstars.
Below, Ahmed (DH) chat with agfunder news (AFN) About the future of indoor farming, the characteristics of hexafarms, and why sheep and AI don't necessarily mix.
AFN: What does Hexafarm offer farmers?
D.A.: Cameras and sensors are installed around the farm, providing harvest record data from producers. The system is highly dynamic and can tell growers how much to harvest just from the data collected. [will] Harvest a few weeks in advance. Knowing this, the next step is to provide producers with viable strategies to get the most out of the resources they put into it.
The ultimate “holy grail” is pushing the boundaries of AI and machine learning in agriculture. With all due respect, when we talk about the current state of machine learning in agriculture, it's just a statistical method and it doesn't work.
We're working with 14 different farmers, each with a completely different farming setup, and that's one machine learning model. So the goal is that once you understand the patterns, machines will actually be much better than humans at optimizing production in real time.
AFN:Why did you establish the company?
Da: I was working with farmers from upstate New York who go to farmers markets in Manhattan. Sometimes we would leave with an empty truck, and other times the truck would be full of cabbages. I thought, “This isn't working, I need to do something else.”
I started asking fundamental questions about how we could improve food production, and found that all existing approaches had fundamental problems. On a commercial scale, plant biology is the least optimized thing by growers. If you dig a little deeper, you'll see that there's a lot to optimize.
Nothing fundamental has changed since the 1960s and 1970s, other than the introduction of chemical fertilizers, pesticides, and controlled irrigation. [in commercial food production]. I have made it my mission to start the Fourth Green Revolution by leveraging digital technology at a highly affordable price to meet the needs of millions of plants in real time.
I knew I wasn't the first to think of this. Look at all the greenhouses and indoor farms out there. But we thought we could tweak and adjust things to get more production out of farming. You can see that the producer still has 30% left.
[The company] We took the Techstars Berlin program. Then we went to the greenhouse to see what an ideal market looks like. We currently have more customers than we can handle. By the end of 2024, our systems will process approximately 25 million kilograms of tomatoes and strawberries.
AFN: How does AI in your technology benefit indoor farmers and others?
Da: Suppose you are growing tomatoes. What we currently offer is a short-term perspective, such as forecasting products that are already selling like hotcakes.
This is a typical case for my client. They had 5 million euros before the season started and needed to grow X amount of tomatoes to supply the supermarkets. So now that they had the money, they had to produce tomatoes. They need to know what production will actually look like, sometimes within a margin of hundreds of kilograms. There are three to four head gardeners on the farm who examine and forecast production. They have to count fruits and flowers and many invisible things. But they are never correct.
Next comes Hexafarm. We have these cameras and some sensors, and the model basically looks at increases and decreases in the number of fruits and flowers. We look at historical data, i.e. the transparency of greenhouses, including how ultraviolet light acts.
We don't hardcode these rules either. We don't set things up like, “We need to check these five variables.” We're like, “Here we have a bunch of data, here we have human experts and historical records.” Now the algorithm has to come up with a way to understand the work itself.
So far this approach has worked very well. As we collect more data, we can easily exceed the previously known limits.
Computer vision extracts food counts, flower counts, leaf area indices, and more. In terms of model performance and benchmarking, we use literally everything from a computer vision perspective. The system then tells you that you have harvested this much in one day, and this much that you will harvest next week.
Every producer says they get 60% to 90% accuracy from time to time, but I'm like, “What if I said this season it's 90% and next season it's going to be 92%? We're constantly improving. I think that's because I continue to do so. . That's another aspect of AI and machine learning.
So we're trying to treat plants as algorithms, and we need to understand this black box and what elements should be there. And we have a lot of humans involved in the loop, and we have a series of AI machine learning processes that rely heavily on computer vision, but that becomes the raw data for predictive models.
AFN: What are the characteristics of Hexafarm?
Da: One unique difference between hexafarms and others is that it requires one large model, similar to what current AI models such as ChatGPT do. The idea is that the model is deep and rich enough to understand the properties. Select for each farm and match it.
As an example, we trained a model on two or three strawberry varieties. Currently, there are about 16 varieties of strawberries in the system. And the actual performance is pretty good. It takes about a week for us to grow a new variety and reach human-level or higher performance.
Adding one more crop is no big deal. Of course, if you go from strawberries and tomatoes to raspberries, it's a little different because you have different dependents, but we're really working towards this new round of funding right now.
AFN: Why indoor farming?
D.A.: We try to stick to CEA. The reason is that these businesses need to really improve their margins and CEA provides the perfect venue for our technology. It's more concentrated, yields more returns, and has less variability, which makes tools like ours easier for indoor growers to use.
But nothing fundamentally prevents us from going out, and in fact, we do it.
We also service foil tunnels and plan to expand into regular fields once our AI models understand the crop a little better and it becomes easier to obtain clean data. We recently set up cameras in the field and used the data to check the number of sheep.
We have customers in Germany, the Netherlands, Austria and Switzerland. You can see cultural differences throughout the greenhouse.
Some are very elegant, while others won't actually allow you to step into the greenhouse. They don't trust anyone, so we have to let them install the equipment. Other companies tend to produce as much as possible, doing things like moving table tops.
When I was raising money, an investor pointed out that the market was fragmented and asked how we would respond to that. I love software. I can write something that almost anyone can use. And I love this fragmentation. That really shaped our tools.
AFN: Some areas of indoor agriculture have gotten a bad rap over the last few years. What are your thoughts on indoor agriculture as a whole?
D.A.: It is 100% possible to grow food profitably using an indoor farm. But some investors just haven't read the history books, and companies were using systems that didn't even have something as simple as a CO2 sensor installed on their farms.
I can make predictions about the greenhouse market and CEA in general. It will continue to grow. Climate change is real, and so is the demand for high-quality, healthy produce. All of our greenhouse customers are already profitable, but we want to make them even more efficient and at the same time give their fledgling new business a head start, whether you start your greenhouse in the Netherlands or in Dubai. I'm thinking of letting you do that. .
