Business Problem & Solution

Guest reviews are the marketplace's load-bearing trust signal — they decide which properties get booked. But raw submissions are noisy: promotional spam, leaked phone numbers and emails, personal abuse, and low-effort one-liners all arrive in the same stream as the genuine, useful reviews. Publish them unfiltered and the trust signal degrades for the whole ecosystem.

Review & Content Quality Intelligence is the gate between a submitted review and a published one. It classifies every review, scores its quality, applies an owner-controlled rule overlay, and routes each to approve, hold, or reject — so only clean, useful reviews reach the property detail pages that the rest of the apps read.

Pure rules are predictable but brittle; a pure classifier is flexible but opaque. This app uses both, in layers the owner can reason about: an immutable Layer-1 baseline (contact-info, promo-spam, low-effort repetition) that can be extended but never weakened, and a Layer-2 owner overlay (blocked phrases plus a hold-below-quality threshold) the operator tunes in the console. The metered classifier sits on top, producing a category, a toxicity score, and a quality score — but the owner always makes the final call.

The demo proves the gate end-to-end: an approved review is upserted into the shared core review store and renders on the property detail page, with no second copy of the truth. It simplifies the model itself — in the prototype the classifier is a deterministic, content-shaped simulation, and Stage-2 swaps that body for the real dual-mode call behind the same signature.

Synthetic reviews only — invented guest names, titles, and bodies over the shared synthetic property catalog. No real guests, no PII, no real brands. Categories, toxicity, and quality figures shown anywhere are representative and labelled, and the seeded owner overlay (a small blocked-phrase list, hold-below-40) is a default the owner can refine.