Plant · live
Spotify Insight
A team of AI agents read all 660,622 Spotify app reviews and argued their way to one product priority. Every number on the dashboard traces back to evidence.
- 660,622reviews, none missing
- $7.94for the full run
- 94%topic accuracy on a hand-labelled set
How it works
- An enricher agent (Claude Haiku, in batches) classifies every review by topic, intent, severity and free versus paid tier.
- A blind verifier agent re-checks a 1,000-review sample. It caught 20 of 20 planted errors.
- A grouper builds the issue taxonomy, and a diagnosis agent maps every complaint to one of 31 concrete symptoms.
- A memo writer drafts the recommendation, and code checks every number and quote before a critic agent signs off.
- The verdict: fix playback first. Crashes and music that stops on its own lead every lens.
- The dashboard reads from Postgres and makes no model calls, so it’s cheap to keep online.
Built for Fundamentals of Agentic AI, Berkeley Haas, Fall 2026.