Sensory evaluation platform

The problem

Sensory analysis is how a food manufacturer answers "can anyone actually tell the difference": between an old recipe and a new one, or between two suppliers of the same ingredient. It is a laboratory discipline with published methods, and the statistics are the whole point. Whether a panel's result is significant depends on the panel size and the chosen significance level, read from published critical-value tables.

What I did

Two years of continuous feature and correctness work across the API and the React front end: methodology configuration (which options each method exposes, how discriminants and product categories bind to methods), scale handling for the bipolar and five-point methods, the results page, company and respondent management, and a long tail of fixes in the places that matter.

That class of bug is the reason the work matters. A wrong boolean in methodology configuration does not crash anything; it silently changes which questions a panellist is asked. A significance threshold wrong by one produces confident, plausible, false conclusions about a product going to market.

A decision and its cost

Panellists ranked samples by dragging them into order. Then the panel moved to a kiosk setup where drag-and-drop was not reliably usable, and I replaced it with explicit move-up and move-down controls. The faster interaction on a normal device was lost; in return it works on every device the actual users have, and it is keyboard- and screen-reader-operable by construction. The input device is part of the requirements, and it changes after you ship.