Every week, leadership at a corn processing facility asked the same question: what will the plant produce this week? Nobody could answer with confidence.
The plant turned corn into three core outputs: starch, oil, and gluten. Yields moved in ways nobody could fully explain. Corn quality, oil content, and hundreds of other process variables were in play, and no one knew which ones actually moved output. Operators adjusted based on experience and instinct. Sometimes it worked. When it didn't, no one could say why.
The plant had plenty of data and plenty of operator experience. It lacked a framework to connect them, one that could tell operators which levers mattered and what would happen when they pulled them.
The objective was concrete: improve site yield and make it predictable.
Falconi's team started on the plant floor, mapping the production process end to end and ranking the variables most likely to drive yield. Then came the data work: pulling the right historical records and structuring them into a database the models could learn from.
From there, the team built a hybrid ensemble of two machine learning models, a Support Vector Machine and an Artificial Neural Network, trained on the plant's own history. The models learned how inputs like corn quality and oil content shaped what came out the other end.
A model sitting on a data scientist's laptop doesn't change how a plant runs. So the final phase turned the analytics into tools operators could use: optimal operating points for each production scenario, site-specific playbooks, a rollout plan for other sites, and a governance structure to keep the models honest over time.
The oil yield model explained 79% of variance in the plant's most valuable output. Weekly predictions landed within 0.17 percentage points of actual yield. For the first time, the plant could see what was driving its results before the week happened, not after.
Running at the model's recommended operating points, the projected gains added up:
The model also flagged a trade-off: pushing more corn toward high-value outputs meant producing less bran, a lower-value byproduct. Even after that hit, the estimated net gain came to $1.74M a year.
Fractions of a percent, multiplied across everything the plant produces, all year long.
The variable prioritization happened before any modeling, so the analytics answered questions that mattered on the floor. The models simulated future scenarios rather than explaining past ones, which gave operators yield projections for each production setup. And the engagement ended with playbooks, operating points, and governance in the operators' hands.The models only pay off when the people running the plant can act on them. That's where Falconi spends its time.