News
Kubernetes on a tractor: how precision farming is revolutionizing agriculture
From complete failure to 40 customers in two years. The Aurea Imaging story shows that even the most challenging edge computing projects can be tamed.
At Edgecase 2025, Wieneke Keller (CTO) and Sebastian Lenartowicz (Software Engineer) from Aurea Imaging shared a story everyone recognizes: a project that completely fails but eventually grows into a successful solution. Their TreeScout device brings Kubernetes to a place it's never been before: on tractors in orchards.
From apples to algorithms
"Want to become an apple grower in five minutes?" Wieneke began their presentation. She asks the crowd which of the three apples is the best apple, showing three differently sized apples on screen. Behind that question lies a complex challenge. An apple of the right size brings in 20 cents more. For an apple grower with thousands of trees, that can mean the difference between profit and loss.
Next question: "Which tree is best?" Here's where the science gets interesting. A tree gets all its energy from the sun, water, and soil, then spends that energy on something. Too much blossom means the tree produces many small apples. Too little blossom results in fewer but larger apples. The sweet spot is finding the optimal number of blossom clusters per tree to ensure apples reach that profitable size.
That's what precision agriculture is all about: treating each tree individually to achieve maximum yield. Their device, TreeScout, scans all trees in an orchard, detects how much blossom is on each tree, and gives the farmer a map to apply targeted blossom thinning.
Then everything went wrong
The first season in 2023 was nothing short of dramatic. "I called our driver, our field engineer, and asked what he'd been drinking the night before, because his GPS track looked pretty zigzag," Wieneke recalled. The driver hadn't been drinking at all; the system just didn't work.
Of the eight customers, only six delivered usable maps. Farmers had to wait a month for results, while blossoms only stay on trees for two weeks. Support was in panic, and developers had no clue how to solve the problems.
The cloud native choice
Sebastian picked up the story: "We faced a choice. Would we take the conventional route with Robot Operating System and C++, or gamble on something that was virtually untested at the time: Kubernetes at the Edge?"
The numbers spoke for themselves. With ROS, it would take 12 to 18 months to get back to where they were. With K3s, they managed it in four months. "We're not a big team: 25 people, of which eight are software engineers. We played to our strengths: cloud native, Python, and machine learning."
Challenges at the edge
Building a device for tractors brings unique problems. "Imagine if the only way to shut down your device is a hard power cut," Sebastian explained. "Think about what that means for your software."
You also have to work within ecosystems dating back to the 1990s. Standards for tractor peripherals and high-precision GPS from the maritime industry: all legacy systems you need to interface with.
And then there's NVIDIA. "Cuda in the cloud is bad enough, but try Cuda on specialized edge hardware. It's a loose collection of half-baked science projects." Still, it can be tamed if you put in enough work.
Success through jumping together
The second season went much better. Time from scan to insight: from a month to an hour. From eight to ten customers, and everything worked. Last season, 40 devices were already driving around, though that's starting to create new challenges.
"One of our most important lessons," Wieneke reflected, "was to jump together. We could only test blossoms when there actually were blossoms, so we flew to South Africa. We found a hundred bugs, but it was a test. Then we fixed everything."
Customers proved surprisingly patient with incidents, as long as you solve them quickly and pick up the phone. "Everyone on our team has literally had their feet in the mud solving problems."
Edge computing comes of age
The TreeScout story from Aurea Imaging shows how far edge computing has advanced. From a Python monolith in a single container, they've grown to eight true microservices, all running on K3s. On a tractor.
Sebastian concluded with an important lesson: "Expect the unexpected. What can go wrong will go wrong. You can't test everything from the office. You have to go to the field."
And sometimes your presentation crashes, but then you just keep going. Just like precision farming, it's about persevering, learning from mistakes, and tackling challenges together.
TreeScout generates about 10 million data points per device during a two-week blossom season. For a startup of 25 people, that means getting creative with scalability, and proves that cloud native architectures can work in even the most challenging environments.
