Photo in, grade out: building an agri marketplace grading agent

The problem

Buyers on a B2B commodity marketplace need to trust a grade they cannot inspect in person. Manual grading was slow, inconsistent between graders and impossible to scale across lots arriving from hundreds of sellers. A wrong grade in either direction has a price, and the two prices are not the same.

• Over-grade and the buyer receives worse than they paid for. Trust in the marketplace is gone.

• Under-grade and the seller is short-changed. Recoverable with a re-grade.

• So the rule agreed on day one: under-grade cautiously, never over-grade.

What we built

A vision model estimates grade, defect rate and foreign matter from a photo or short video. Anything below the confidence threshold goes to a human grader with the model's estimate attached. The threshold is set from the two error prices, not from a round number. Alongside it, a price forecasting engine gives short, near and long horizon forecasts with confidence bands.

Every disputed grade becomes a new case in the evaluation set. The set has grown with the marketplace.

Every disputed grade becomes a new case in the evaluation set. The set has grown with the marketplace.

Every disputed grade becomes a new case in the evaluation set. The set has grown with the marketplace.

What broke first

Lighting. Phone photos taken in a warehouse at dusk looked nothing like the training set. The fix was not a better model, it was a capture guide for sellers and a confidence penalty for low-light images so they route to a human instead of a guess.

What runs today

Both systems are live on the agri marketplace with escrow-backed settlement built with banking and trustee partners. Miss rate and over-grade rate are reported separately every week. The over-grade line has stayed where the rule says it must.

If you have a judgement call that a person makes hundreds of times a day from a photo, a document or a screen, it is probably a candidate. Tell us what it is.