Real Vet Reviews

What it is

One page per clinic, built from every rating that exists for it — the Google and Yelp aggregates, Facebook recommendations, state license and complaint records, and reviews from members who actually paid the bill. Every number on the site comes from a real rating or a real receipt. Nothing is invented, and a clinic with fewer than ten ratings gets no score at all.

The data pipeline runs on a schedule: discover clinics for a tracked location, fetch only the reviews newer than the last run, have a local LLM tag each one by topic and sentiment, rescore, and write an AI summary that is labeled as AI. The scoring method is pure, unit-tested, and explained to readers on its own page.

The whole thing runs on a 2 GB container I own. The dev machine compiles and tests; the server only runs. If a release fails its health check the previous image is restored automatically.

“Clinics with fewer than 10 ratings get no score. Nothing is invented.”Real Vet Score module, src/lib/score.ts

The parts that were hard

• Merging one clinic's Google and Yelp listings into a single profile — by phone first, then name plus ZIP — without ever double-counting a review
• A Bayesian score that a handful of reviews can't dominate: platform aggregates plus member reviews weighted 1.5×, or 3× with a moderator-checked bill, shrunk toward the all-clinic average with the weight of twenty reviews
• Serious allegations never move a score alone — three or more moderator-confirmed within 24 months do, scaled by how far the clinic's rate exceeds the norm
• Incremental daily refresh against paid scraping APIs, with usage and balance tracked so a bad run can't drain the budget
• LLM topic and sentiment tagging on a local model, so review text never leaves the box
• Nightly pg_dump and object-store backups to the NAS, with a restore actually tested