01 — Case study · Finary · Jupiter
App Review Intelligence
Classification pipeline turning 1,965 App Store & Play Store reviews into a live product-intelligence dashboard the team monitors at scale.
Year
2026
Role
Growth Data & Automation — solo DRI
Context
Finary
Stack
Python · Claude Haiku · Streamlit · Pandas
1,965
reviews classified
2
app stores covered
1
live dashboard for the product team
The challenge
Finary's product team had thousands of App Store and Play Store reviews — raw, unstructured, in several languages — and no scalable way to know what users were actually complaining about, or whether an issue was growing. Reading them manually doesn't scale; ignoring them loses signal.
What I built
- A Python ingestion pipeline collecting and normalizing 1,965 reviews across both stores.
- An LLM classification layer on Claude Haiku tagging every review by theme, sentiment and severity — prompts iterated against a hand-labeled validation sample until the taxonomy held.
- A Streamlit dashboard in Finary's dark branding: theme trends over time, store and version filters, drill-down to the raw verbatims.
- A design meant to re-run on every new batch of reviews, so monitoring is continuous rather than a one-off study.
Results
- The product team got a scalable review-monitoring system instead of anecdotal reading.
- Classification surfaced sync reliability and pricing perception as the dominant recurring themes — concrete, prioritizable levers.
- The pipeline pattern (ingest → LLM classify → dashboard) became reusable for other feedback sources.