App Store reviews

The testClassification at volume: a hundred real reviews, five questions each, in a few seconds and for a fraction of a cent. Messy text, typos, mixed languages. Does the model hold up, and what does it cost to run this every day?

How to read itThe feed is the app's most recent reviews, unfiltered: praise, complaints, everything. The big number is severity, 0 to 3; the default order is worst first, so praise sinks to the bottom. Flip the sort to Newest or Rating to see the rest. The red tag is the review type with the model's confidence; area comes with its probability; "blames update" is a yes/no probability. The rails on the left count how the last hundred reviews split.

What to watchp50 latency and total cost in the strip. Whether "complaint" vs "bug" is a distinction the model actually makes. Language detection on short reviews. Try a store in another country and compare.