From Greenhouse Data to Real-World AI Systems

We started a hydroponic greenhouse in Asunción in 2018. Not as a research project. As a business. From day one we supplied Bolsi — a restaurant that would later rank among the top 20 in the world. That set the bar immediately. You don’t deliver to a kitchen like that and guess, so we didn’t.

From the first cycle, we were logging everything: temperature, humidity, nutrient concentration, light. Not because we had an AI thesis — because if a crop failed, we needed to know exactly why. We built our own sensors. Printed enclosures. Wrote scripts to log and visualize the data. Iterated constantly. Standard engineering process — just applied to lettuce.

Three years in, we added computer vision and machine learning. The impact wasn’t theoretical. It was operational. Tasks that took hours of manual inspection became seconds on an edge device. Issues we used to approximate became measurable. Systems that depended on someone being physically present no longer did.

That changed how we thought about the work. We started asking a simple question: what else around us is slow, manual, or dependent on someone being there at the right moment?

We looked at the people we were already working with — researchers, field technicians, local agencies. The same pattern kept showing up. Critical data being collected by hand, on paper, by people standing in a river or driving a truck down a dirt road. Paraguay’s rivers have migratory fish counted manually by technicians on boats. The Gran Chaco has camera traps that sit for months before anyone reviews the footage. Municipalities have no systematic way to know the condition of their own roads unless someone drives every kilometer and reports back.

We’re currently building several systems around these problems. We’re entering them into Moonshot 2026 and will be launching them in coordination with the event dates — more on that as things become public.

These aren’t edge cases. They’re critical systems — just invisible to global AI platforms because the data doesn’t exist and the market doesn’t fit. So we build for them. Fish detection pipelines in the Paraná basin, with humans in the loop. Trail camera analysis for the Chaco, built with local scientists. No pivot deck. No market sizing exercise. Just years of building systems in the field, in a place most platforms ignore. We’re FDF Labs, based in Asunción. We build AI that works where data is scarce, conditions are messy, and reliability matters. If you’re working on problems like that — we should talk.