AI sentiment analysis

Free-text answers are the most valuable thing a survey collects and the least likely to be read, because a few hundred of them is more than anybody works through on a Tuesday afternoon. Classification is what makes that pile navigable.

Every written answer is classified as positive, neutral or negative, and the response feed filters on it. The negative bucket is the one that earns the feature: it is where the specific, actionable complaints are, and it is exactly the set somebody scrolling a chronological list will never reach.

The classification is per answer rather than an aggregate mood, so the score and the sentiment can be read against each other. A high score with negative comments is a real and interesting disagreement, and one that averages would hide completely.

It is honest about what it could not do. An answer the classifier could not use is marked as processed with no classification rather than left looking like it is still in a queue, so the count of unclassified replies goes down rather than showing a number that never moves. Answer lengths are capped before anything is stored, since the text ends up in a model prompt.