The farmer is 52. He inherited the place from his father and turned it into an agritourism operation during the pandemic — guests hike, eat cheese that is made downstairs from the dining hall, and watch the sun set behind the ridge. The cows are part of the product. They also keep walking into his neighbor's corn and his neighbor is losing his patience.
Three years ago he bought GPS collars. They work, mostly. When it rains, the pings get sparse. When the cows walk down the gully behind the ridge on the east side of the field, the signal loses line-of-sight to the cell tower and he gets nothing for an hour or two. A cow can cover four kilometers in an hour. By the time the signal comes back, she is nowhere near where the collar said she was.
On a Wednesday in early March, one of them went missing in a storm. The farmer walked the fence line with a flashlight from 9 pm to 1 am. He found her the next morning a kilometer down the road, blocking a tractor. His neighbor was not amused.
That night, back at the house, he opened his laptop. He searched for what he thought would work. His exact query — we know this because we asked him later, and he showed us — was:
find cow far field gps problem
Google gave him Reddit threads about cattle GPS reliability. YouTube gave him advertisements. One of the sidebar links, eventually, pointed him to DRONA. He almost didn't click. He doesn't know what a drone is, not really. He knows they are expensive, they crash, and they are toys for his nephews.
He clicked anyway. And typed the same thing again.
The conversation
The chat did not give him a drone. It asked him a question. How many cows, how big is the field, and when do the cows go missing — night, rain, or both?
He typed: "40. 60 hectares. Mostly in rain and at dusk."
That was enough. Behind the scenes, the industry-classifier routed the query through the agritourism long-tail branch into mission template AG-03 (livestock locate) with environment tag rain and mission window dusk. The form-factor recommender flagged multirotor hex over quadx on the basis of wind tolerance and flight time. The compatibility engine enforced the IP54 rule as soon as the rain tag landed. A thermal radiometric payload was selected because cattle body heat against cold-soaked pasture at dawn and dusk is a 6–10°C contrast, which is trivially resolvable on a FLIR Vue TZ20 even at 80 m AGL.
The engine computed all-up weight (4,150 g), thrust (12.4 kg), TWR (2.99), flight time (27.5 min), and cost ($2,380 parts). It ran the 12 universal rules plus the agriculture-specific predicates. It passed.
He saw a 3D drone assembling itself on the right side of his screen. He sat there for what he later described as "four or five minutes" without typing anything. Then the chat asked him whether he wanted to see the parts list.
herd-watch-01 · generated configuration
The design the chat produced. Every line was computed against the 12-rule compatibility engine and the agriculture-specific predicates.
motors · T-Motor MN3110 · 470 KV · ×6
battery · 6S 6000 mAh · 30C
fc · Pixhawk 6C · ArduPilot Copter
thermal · FLIR Vue TZ20-R · radiometric
visible · Sony IMX678 · 48 MP
autonomy · L3 · scheduled 5:15 · 11:30 · 17:45 · 22:00
hover time · 27.5 min · AUW 4,150 g · $2,380
Why it worked
There are probably a dozen reasons, and we think honesty requires us to name them all — including the uncomfortable ones. The farmer is not a pilot. He is not an engineer. He did not know what an ESC was before that night, and he does not need to know today. Most of the reasons DRONA worked for this farmer have nothing to do with him. They have to do with not making him feel stupid for not knowing, and with the tool doing, without asking, every piece of judgment he could not have done himself.
The chat never made him name a drone. It never asked him to pick a form factor. It did not ask him to reason about thrust-to-weight, or compatibility, or which flight controller. It asked him three things: how many cows, how big, and when does it rain. Every other decision the tool made, silently, and then showed him.
The physics mattered in ways he will never notice. The rain tag enforced IP54 — if the chat had proposed a non-IP-rated frame, the compatibility rule would have caught it. The thermal payload was chosen because the mission tag livestock classifies into the thermal category. The autonomy level was set to 3 because a 60-hectare scheduled search is beyond manual-pilot range, and also beyond his certification, but the ArduPilot waypoint flow is within his training envelope.
The timeline
21:42
The farmer opens the chat, types nine words, answers three questions. Nine minutes in, he has a complete design — configured, priced, 3D-rendered, compatibility-verified — sitting in front of him.
His nephew, 19, a mechatronics student in Tirana, recognizes the Pixhawk 6C and T-Motor MN3110. "Uncle, this is serious. Not a toy." They order the BOM together from a reseller in Tirana.
Six motors, the X500 frame, the Pixhawk, the FLIR Vue, the Sony sensor, the props, the wiring, the Holybro GPS module. Total: $2,420 including shipping. Close enough to the $2,380 estimate.
Nephew and uncle assemble over two Saturdays. They use the Markdown build guide DRONA generated alongside the BOM. They get stuck once — on the UART allocation for the thermal-camera serial — and the chat walks them through it.
The .param file loads into Mission Planner on the first try. The autonomous waypoint mission is generated from the 60-ha polygon the farmer draws on Google Earth.
A still morning, 15 m AGL, a manual takeoff and hover test. It flies. The thermal feed shows two cows in the lower field. He tears up. So does his nephew, honestly.
5:15 am. Drone launches from its box on the barn roof, flies the preset ridge-run route, returns home in 23 minutes with thermal imagery of 38 of 40 cows. The missing two are later found to have wandered into the neighboring olive grove. The neighbor is not happy. But at least the farmer is the one who tells him first.
Four scheduled flights per day. Guest count at the agritourism up 12% over the same period last year — the drone is part of the visit now. No cow has been lost for more than an hour since the first scheduled mission.
The numbers
What this is not
We do not want to tell this story as if it ends every time. Plenty of first-time builders do not have a mechatronics-student nephew. Plenty of rural agritourism operators cannot get a Pixhawk delivered in a week. Plenty of chats do not land on a design that works on the first flight. We are sharing this one because the conversation is archetypal of the motion we want the product to support, not because every session ends this way. Most sessions do end with a renderable, compat-verified design — we measure that — but the gap between a rendered design and a working machine is not always as short as it was in this case.
We also do not want to overclaim what the AI did. The AI did not design a drone. The AI asked three questions, classified intent, and invoked a deterministic physics engine and a parts catalog that we maintain. The intelligence is in the grounding, not in the language model. That is a deliberate architectural decision — documented in our physics primer — and it is the one most responsible for the fact that this farmer's drone exists and flies.
That is exactly the sentence we would write on the front page of the website if the board would let us.
This case study is lightly anonymized at the farmer's request. The farm, the herd count, the chat transcript, the build timeline, and the price are real. The name has been removed entirely.