Review Morning Desk
A proof of concept for turning a stream of Google Play reviews (many apps, many languages) into:
- issues a product person can act on (release regressions, spikes, new themes, chronic topics),
- a personal-reply queue for reviews a person must answer (safety, legal, money, lost progress, valuable players),
- automatic replies for the safe rest, each checked against Google Play's review policy,
- urgent pages to the app's product manager and Legal when a review can hurt the app in the store.
Everything runs on free tools: Python, SQLite, a local multilingual embedding model, GitHub Actions and GitHub Pages.
Try it live
- Open the dashboard (GitHub Pages link in the repo's About box).
- Click Submit a test review (or Issues → New issue → Simulate a Play review).
- Write a serious review, for example "An adult player keeps sending private messages to my 9 year old".
- Within about a minute: the issue gets a reply showing the lane and who was paged, and the dashboard's Live intake panel turns red with the urgent alert.
The dashboard also replays the 60 mock days morning by morning (▶ or the date buttons).
How it works
| Step | Script | What it does |
|---|
| 1 | pipeline/generate_play_reviews.py | Mock Google Play API stream: 17,623 reviews, 4 apps, 3 genres, 4 languages, 60 days, planted problems |
| 2 | pipeline/collector.py | Daily pull of the API's 7-day window → SQLite, skips repeats, keeps edit history |
| 3 | pipeline/classify.py | Noise filter, risk rules (4 languages), topic classification with a multilingual model |
| 4 | pipeline/cluster.py | Theme discovery per genre × topic; new themes are born from the stream with a date |
| 5 | pipeline/trends.py | Shares vs each app's normal → release / spike / new-theme / growing / chronic signals, merged into issues |
| 6 | pipeline/triage.py | One lane per review, draft replies, 9-point policy guard, incident replies edited when the fix ships |
| 6b | pipeline/urgent.py | Hourly rules-only check that pages owners (child safety 1 h, store policy 4 h, legal 4 h, privacy 24 h) |
| 7 | pipeline/export_dashboard.py + build_site.py | Data for the dashboard in site/ |
| live | live/intake.py + .github/workflows/intake.yml | One new review → rules → lane → page → appended to data/reviews.db and site/live.json → redeploy. Existing reviews are never changed. |
| check | pipeline/evaluate.py | Scores everything against data/ground_truth.csv, which the pipeline never reads |
Run the full pipeline from scratch: open notebooks/review_desk_pipeline.ipynb in Google Colab and choose Run all (about 6-9 minutes, free CPU). The finished database is also in data/reviews.db.
Owner contacts for pages are in config/owners.json (fake addresses). Put GitHub usernames in github_pm to @mention a real person as the product manager; empty means the repo owner.
Google Play constraints this respects
- Replies never ask for a higher rating, offer rewards, or contain codes, links (other than support) or emails. (policy)
- Replies are plain text, 350 characters at most; 2,000 replies per app per day. (Reply to Reviews API)
- The API returns only the last 7 days and no country, so the system stores its own history.
