Strawberry harvesting in commercial agriculture is one of the few remaining tasks that has resisted mechanization at scale. The fruit is delicate, the cutline between ripe and underripe is narrow, and labor costs have become a structural problem for the industry. When the University of Maryland's Bio-Imaging and Machine Vision Lab took on the problem, the goal was a computer-vision system capable of processing strawberries at commercial throughput with selectivity comparable to skilled human harvesters.
I led product development from initial design through commercial deployment. The resulting system processes 16 tons per hour, recovers 15% more fruit than the existing manual process, and saves thousands of labor hours per facility each week. A partner organization purchased the rights to the platform and commercially deployed a revised version.
The constraints
The product had to work in an environment most computer-vision systems aren't designed for: direct sunlight, dust, vibration, variable fruit density, and an operator profile of agricultural workers — often non-technical, using the system across long shifts in challenging conditions. Touchscreens didn't work, because workers' hands were often wet or gloved. The fruit moved through the system at rates that ruled out human supervision of individual harvesting decisions.
The product wasn't just the computer-vision model. It was the full system: the optical hardware, the conveyor mechanics, the cutting and sorting actuators, the interface designed for the actual operating environment, and the deployment process for getting the system installed and operational on partner farms.
What was built
A modular hardware platform with a machine-learning vision system trained on cutline detection across strawberry variants and ripeness states. The model identified harvest-ready fruit with selectivity comparable to skilled human harvesters, while operating at throughput rates no human could match.
The product was designed around the deployment environment from the start. The interface was optimized for wet-gloved operation. The mechanical design used IP68 environmental certification for outdoor use. The system was modularized so components could be serviced or replaced in the field without specialized equipment. We spent six months on-site at commercial facilities in California during commissioning, refining the system against real operating conditions rather than lab simulations.
What it shipped
The platform processed 16 tons per hour in commercial operation. Against the manual process it replaced, it recovered roughly 15% more fruit yield — the model's cutline selectivity was more consistent than human harvesters over long shifts. Labor savings ran to thousands of hours per facility each week. A partner organization purchased the rights to the platform and deployed a revised version into broader commercial production.
What this case shows about how I work
The work demonstrates several things at once: 0→1 product development on hardware with real-world deployment constraints; applied machine learning where the model's output controls a physical system rather than a screen interaction; design for the actual user environment rather than the abstracted user; and a willingness to spend long deployment cycles on-site refining a product against real conditions. The through-line is the same as in the software products I've shipped since: customer discovery in the operating environment, not the conference room; architectural decisions that account for the operating constraints, not the ideal case; and a willingness to commit long deployment cycles to getting the system right in the environment that matters.