Worked example Workflow mapping: feed formulation

Least-cost feed formulation is a solved problem. Trusting the nutrient matrix is not.

Every feed mill re-solves the same linear program: meet the nutrient spec at minimum ingredient cost. The optimization is decades old and it is not the hard part. The hard part is the data underneath it. When maize moves, when a new soybean-meal lot lands, when the assay disagrees with the book value, the answer is only as good as the matrix it ran on.

August 14, 2026

See if the Sprint fits

Abstract formulation pipeline: a grain kernel feeding an optimization network that resolves into a ration list on a mobile screen.

The workflow, mapped.

A formulation run touches four data surfaces: the ingredient price sheet, the nutrient matrix (book values plus any in-house assays), the species requirement spec, and the constraint set (inclusion limits, ratios, regulatory caps). In most operations those live in a spreadsheet, a binder, an instrument export, and someone's memory, and they converge by hand every time prices move. The formulation software is not the bottleneck. The assembly and the trust in the inputs are.

The worked example.

I built a small demonstration pipeline on synthetic data: a synthetic price sheet for a dozen common ingredients, a nutrient matrix where every value carries its source tag (book value, synthetic assay, or flagged unknown), and a broiler-grower spec. The linear program solves in milliseconds. The useful part is what surrounds it: every nutrient value in the solution traces to its source tag, any value that cannot be traced is flagged rather than silently defaulted, and a re-price run shows exactly which matrix entries drove the ration change. Cited or flagged, applied to formulation.

What transfers

The method, not the mill.

The optimizer was never the moat.

Any competent tool minimizes cost against a spec. The differentiating asset is a nutrient matrix your nutritionist actually trusts, with provenance on every number.

Re-pricing is a workflow, not an event.

When a commodity moves, the question is not "re-run the LP" but "which inputs changed, who verified them, and what did the change do to the ration." That is a pipeline with a review step, not a spreadsheet race.

The assay loop is where the value hides.

In-house assay data (NIR or wet chemistry) is worth nothing sitting in instrument exports. Piped into the matrix with a source tag and a review gate, it compounds on every formulation run.

What this is, and what it is not.

This is a worked example on synthetic data. It is not a commercial formulation product, and no client result is claimed. It demonstrates the same method as every other example on this site: map the workflow, keep the deterministic parts deterministic, put provenance on every number, and keep the domain expert in the decisions that matter. My background in this domain is on the record: named-inventor US patents in metabolic engineering, and career work in animal-nutrition analytics and diagnostics described in the career record.