Smart
Greenhouse
At the center of our work is a 650 m² facility in Asunción, Paraguay. Built as a commercial greenhouse in 2018, it evolved into a unique fusion of commerce and research — an AI R&D lab that continues producing food at scale.
m² hydroponic facility
heads of lettuce per year
in continuous operation
Commerce and research under one roof
The greenhouse maintains its commercial operation while serving as a controlled testing ground for our research. It allows us to generate valuable data, experiment with automation and refine machine learning models in a real-world environment. Each year we produce close to 100,000 heads of lettuce, using the greenhouse not just as a farm but as a living laboratory.
Each growing cycle is an opportunity to run new experiments — test plant health approaches, improve disease resistance and find how to achieve more with fewer inputs. Early trials already showed drastic reductions in water consumption compared to traditional agriculture, along with more precise nutrient delivery and much less waste.
Tech StackTechnologies pushing the boundaries of agriculture and research
Environmental Controls
Automated systems regulate temperature, humidity and airflow at multiple points across the facility. Spatial gradients reveal cold zones, humidity pockets and uneven airflow.
Root Zone
Nutrient solution temperature, electrical conductivity (EC) and pH per channel. Roots live in a different environment from foliage — and that environment has its own behavior.
Light (PPFD)
Light intensity measured at different crop positions. The actual light distribution rarely matches what the design promises. The data shows where there is too much and where there is too little.
CO&sub2;
Carbon dioxide concentration in greenhouse air. One of the most underestimated factors in commercial hydroponics — with significant variation between day, night and ventilation levels.
Computer Vision & AI
Cameras and models monitor development, detect early signs of stress or disease, and measure growth with scientific precision. Time-series images per channel build a baseline of what is normal — and capture deviations early.
Irrigation & Nutrient Delivery
Precision hydroponic systems adjust nutrient mixes and irrigation cycles per channel, minimizing waste and maximizing growth. Solution volume, frequency and total water consumption recorded continuously.
Electronics that survive the field
From dosing pumps to crop management prototypes, we explore how robotics can reduce labor and improve consistency. But the real challenge is not what to measure — it is how to protect the electronics in the environments where they need to work. An enclosure that survives the greenhouse survives the forest.
Why commercial enclosures fall short
In the greenhouse
Sustained relative humidity above 85%, condensation on cold surfaces, chemical aerosols from nutrients, and daily thermal cycles that fatigue seals. Generic enclosures fail due to connector corrosion, moisture ingress and wiring degradation.
In the field and forest
Direct rain, partial submersion in flood events, ambient temperature up to 45°C in the Chaco, UV radiation that degrades plastics in months, insects that find any opening and vibration from branches and wind. No consumer equipment is designed for this.
What we develop
IP67 sealed enclosures
Designed and 3D-printed locally for each sensor type, with tight tolerances for gasket pressure. Cable entries with certified cable glands. Sealed with chemical-resistant neutral silicone. Waterproof to 1 meter submersion.
Passive thermal management
White or silver enclosures to reduce solar gain. Thermal mass design to smooth peaks. For sensors requiring airflow — such as CO&sub2; and ambient temperature — internal labyrinths that allow ventilation without water or insect ingress.
Field connectors
Industrial IP67 connectors at all access points. UV-stabilized cable sheathing. Cable lengths sized for each mounting point. Modular design: the same connector in the greenhouse and in the forest, so equipment is interchangeable.
Edge computing in enclosure
Local processing nodes — Raspberry Pi, Jetson Nano, microcontrollers — go inside the same sealed enclosures. On-device inference and storage. No connectivity dependency to operate: data syncs when signal is available.
Autonomous power
Small solar panels with backup battery for deployments without grid access. Energy management with low-power modes and task prioritization. Minimum design autonomy of 72 hours without solar charging.
Field anti-tamper design
Security fasteners at access points for remote deployments. No external markings identifying contents or equipment value. Low physical profile to reduce visibility. Opening logs via internal sensor.
Mounting methods
In the greenhouse
Perimeter aluminum rails for environmental sensors, adjustable without tools. Ceiling-mounted cameras with single-axis articulation to cover each channel. Submersible probes in NFT channels with stainless steel clamps. All mounts with easy access for maintenance and replacement.
In forest and field
UV-stabilized nylon straps for tree mounting without damaging bark. Steel stakes for ground nodes. Articulated arm to orient solar panels north without clearing vegetation. Fully removable mounting — we leave no permanent infrastructure in reserve areas.
What we learn from the data
The value is not in measuring — it is in understanding which combination of variables predicts final yield. After years of instrumented cycles, we are beginning to distinguish the factors that move the needle from those that are noise.
Root zone temperature has more impact on growth rate than air temperature. Small pH variations sustained over time produce effects that are not visible until week four. The channels on the north edge of the facility consistently produce slower in July — due to thermal gradient, not nutrients.
That kind of knowledge does not appear in any manual. It appears in the data, if the data exists.
From data to harvest
Early problem detection
A deviation in growth rate or leaf color, detected in week two, allows intervention before the problem advances. Without continuous monitoring, those changes are not visible until they have already cost yield.
Harvest window prediction
With enough recorded cycles, the growth model allows projecting several days in advance when each channel will be ready. This simplifies logistics and reduces losses from mistimed harvests.
Future cycle optimization
Each cycle generates data that informs the next one. The conditions that produced the best historical yields become the benchmark for adjusting parameters in the next planting.
Where everything is tested first
The greenhouse is a controlled environment where the cost of failure is low and iteration is fast. Any vision model, sensor integration or data pipeline we develop for other projects passes through here before going out into the field.
A system that does not work on well-lit lettuce has no business in murky river water or in the Chaco forest. The greenhouse sets the minimum bar.
Do you work in precision agriculture or hydroponics?
If you are looking to instrument a facility or understand your crop data, we are interested in the conversation.
Contact Us