Vijay Giri, Graduate Research Assistant, Mechanical Engineering, College of Engineering and Computer Science, University of Michigan – Dearborn
Watch RecordingVijay Giri did not begin his project with a research hypothesis or a model architecture. He began with a conversation with a farmer. A Michigan asparagus grower described the decisions he faces every morning during harvest season — whether to mobilize crews (labor contracts must be locked in days in advance), whether to ship now or hold, and whether a surge of cheap imports from Mexico might collapse the market before a single spear is picked. The harvest window is less than 48 hours. The shelf life post-harvest is three to seven days. The official data resources available to the grower — maintained by the state cooperative and the Michigan Asparagus Advisory Board — hadn’t been updated since 2022. Some cost-of-production figures were nearly a decade old.
In 2023 alone, 650,000 pounds of Michigan asparagus were harvested but never sold — not because the crop failed, but because the market wasn’t there when it entered. A third of the total US asparagus value had collapsed in a single year. Giri and his collaborator — a two-person team with no lab and no funding — decided to build something.
The first challenge was not modeling but data. The market signals that actually drive asparagus price — import volumes, cross-commodity pressures, weather-driven yield shifts across growing regions — are fragmented across sources with different reporting formats, different update frequencies, and no unified pipeline. “Getting them all into one unified pipeline and synchronizing them independently,” Giri said, “was the most challenging part.” More time went into data integration than into training any model. This is a reality that benchmark-focused ML research rarely surfaces: in applied agricultural settings, the hardest problem is assembling the information the model needs to reason about the real world.
The core modeling insight was that yield and price cannot be forecast independently. They are negatively correlated — when harvest is abundant, supply pressure typically drives prices down — and a model that accurately predicts a large harvest without predicting the corresponding price collapse will lead growers to exactly the wrong decision. Giri’s system jointly forecasts both, using a model ensemble including attention-based architectures, combined with a Gaussian copula to preserve the yield-price relationship across the joint distribution. Because no single model architecture consistently dominated across forecast horizons, target variables, and growing regions, the team implemented a performance-weighted ensemble as a safety layer. The output is not a point forecast but a set of Monte Carlo-optimized joint scenarios, from which a profit-maximizing harvest decision is derived.
The practical framing of the results was deliberately grounded: for Michigan growers, the model produces a 2.6x improvement in expected profit and reduces the probability of loss from 44% to 1%. “This model can turn a high-stakes gamble into a sustainable business model,” Giri said. He was transparent about what remained imperfect — the price uncertainty calibration is weaker than the yield calibration — but noted that the joint modeling architecture turns this into a tractable decision problem regardless: given a reliable yield anchor and a full distribution of price scenarios, the question becomes not “what will the price be?” but “what decision maximizes expected profit across that distribution?” The reframing is as important as the model.
Giri closed with a design philosophy that could serve any applied AI project: “We began with the decision. We had a real problem. The model came second. And sometimes the constraint becomes the innovation.”