Observation from the talk: the team deployed, observed the resulting coffee boxes, asked 'is this good?', then adjusted coefficients, added/removed parameters, and redeployed. Over many cycles this loop developed intuition for which parameters were useless versus needed boosting — effectively training the system in reverse, where humans act as the loss function instead of gradient descent. Feels backwards but is a legitimate way to evolve a small scoring system when proper ML is not on the table.