Sales Comment Moderation
A locally trained LLM that moderates comments on sales listings in-house, so customer text never leaves the building to get a label.
Moderation is a classification problem that fails badly in both directions. Miss something and it stays published. Over-flag and you’ve buried real customers’ comments in a review queue nobody has time to work through.
This system moderates comments on sales content with an LLM trained and served locally, with no third-party moderation API anywhere in the path. That was the starting constraint rather than an optimisation. Comment text is customer data, and sending it to an external endpoint just to get a label back is a data flow you have to justify. Keeping the model in-house means you don’t have to.
The other benefit showed up later: running your own model means the policy is yours. Hosted moderation endpoints enforce their vendor’s categories, and those categories almost never match what a particular sales context cares about. Training locally let us draw the boundaries where the business wanted them, then move them when the first version turned out to be wrong in places.
I did the training work and the production deployment.
innoscripta SE · 2026