FeedbackBot: An AI-Powered User Feedback Analysis Platform

I built an AI-powered feedback pipeline that turned a high-volume stream of unstructured seller responses into a classified, searchable, and actionable dataset. By clustering related feedback semantically and routing bugs directly into our issue tracker, the system helped the team spot patterns that manual review had missed and respond to high-value sellers faster.

Before FeedbackBot, every survey response landed in a Slack channel. Someone had to open each one, read it, and decide on the spot whether it meant anything. Slack search only matches exact words, so it worked fine if you already knew what to search for and it did nothing for you if you didn't. Insight lived in whoever happened to be paying attention to the channel, and in whatever word they thought to search for.

Take low stock alerts, as one example. One seller wanted a heads-up before something sold out completely. Another asked why there wasn't a warning when inventory got low. A third just wished the dashboard would flag anything running out. Read one at a time, when you happen to glance at the feed, it might seem like a low-volume request. But in reality the frustration has been building quietly without anyone realizing it.

Triage, clustering, ticket routing, and integration settings — the surfaces sellers' feedback flows through.

Why manual triage stopped working.

Customers use different language to describe the same feature, making search difficult.

At the end of 2025, we released a complete overhaul of our seller admin and began inviting existing users to check it out. Due to a time crunch toward the end of the project, several legacy workflows weren't ported over right away. For four months, feedback about product reordering came in fragments. One seller said they couldn't move products around like the old version. Another asked what happened to drag-and-drop. A third called it unprofessional that their variants displayed out of order after adding a new one. Read one at a time, in a Slack channel full of other feedback, it was easy to miss the pattern.

In total, at least seventeen sellers complained about this feature, but what's even more important is who these sellers were. They were all sellers on our highest tier that had been with us for years, sellers it would be extremely painful to lose over such a silly miss. Fortunately, because the platform surfaced this for the team we were able to quickly restore the functionality and let each seller know personally.

All twelve reports mentioned “packing slip,” but because they arrived over two months, nobody realized they traced back to the same regression.

The problem wasn’t finding the words “packing slip.” Each response described a different symptom—blank pages, cut-off details, overlapping text, or layouts split across multiple pages—and arrived amid a much larger stream of feedback. Read individually, they looked like isolated printing problems.

Once the platform was launched, it quickly grouped them together, we could see that the complaints began after the new admin launched and traced back to a single layout regression. The team was then able to identify the cause and quickly ship a fix.