Yield prediction farmers actually use: lessons from running our own farm
Most agricultural AI dies between the model and the field. Operating Gryn — our own rice, oil, and livestock operation — taught our research center what survives contact with a real growing season.
Agricultural AI has a credibility problem it largely earned. The literature is full of yield-prediction models with impressive validation scores; the fields are full of farmers who have never changed a single decision because of one. When Vyrux Group’s research center began its agriculture track, we made a decision that has shaped everything since: we would run our own operation. Gryn — rice, palm oil, livestock — is where our models meet weather, breakdowns, labor realities, and market prices. The gap between the two worlds is the most instructive thing we have studied.
A prediction is only as useful as the decision it changes
The first lesson was about framing. Farmers and farm managers do not want yield predictions; they want decision support at the moments money is committed — what to plant where, when to apply inputs, when to harvest, what to promise a buyer. A model that says “expect 4.2 tonnes per hectare” is a curiosity. A model that says “this block is tracking below plan early enough to intervene” changes an outcome.
That reframing changes the engineering. Accuracy at season’s end matters less than usefulness at decision points mid-season. Calibrated uncertainty matters more than point estimates, because the question a manager actually asks is “how wrong could this be, and can I afford that?”
The data reality of the smallholder-adjacent world
Most published models assume data environments — dense sensor networks, clean historical records, reliable connectivity — that simply do not exist across most of African agriculture, including parts of our own operation. What exists instead: satellite imagery that is genuinely good and getting better; weather data at workable resolution; and operational records whose quality depends on paper, habit, and the person keeping them.
The research consequence is unfashionable but decisive: invest in the recording layer before the modeling layer. At Gryn, the highest-return “AI investment” we made was making operational data capture — inputs, dates, labor, block-level outcomes — easy enough to survive a busy season. Every model downstream is built on that substrate. Remote sensing then does the spatial heavy lifting that ground sensors would do in richer data environments.
Process beats prediction more often than we expected
The second humbling lesson: some of our largest measured gains came not from prediction at all, but from applying optimization discipline to processing — scheduling around the parboiling and milling bottlenecks in rice, reducing loss between harvest and sale. Post-harvest loss is a data problem wearing an infrastructure costume: you cannot reduce what you do not measure. Before reaching for a neural network, instrument the value chain you already control.
What to do about it
If you are an agribusiness evaluating AI, invert the usual order. Start from the five decisions that most affect your margin, and ask what information would change them, at what accuracy, delivered how — to a manager’s phone, in the language they work in. Fund the data capture that makes those answers possible. Then let models compete for a place in that loop. Our advice comes with mud on it: everything we recommend to clients has either worked on our own farm, or been discarded because it did not.