The problem
Dustiness is a real quality attribute of powder products: how much of it goes airborne when handled. It was being judged by eye.
Judging by eye has two failure modes that compound. It is inconsistent between inspectors, and it is unrecordable: you cannot plot an opinion over time, so you cannot tell whether a process change made things better or worse.
What I built
An automated quality-control system that analyzes video footage of a powder sample being dropped, segments the resulting dust plume, and turns it into a measurement.
The QC team now runs it on 100+ powder samples daily.
Where it landed
The part I am most pleased with is not the model. It is that the measurement became a new quantitative QC standard at the company.
Dustiness used to be a sentence in a report. Now it is a number that can be tracked, compared between batches, set as a threshold, and argued about with evidence. The computer vision was the means; replacing a subjective standard with an objective one was the actual deliverable.