Personalization & Recommendation Platform

A hospitality marketplace lists tens of thousands of campsites and serves guests who book once or twice a year. That combination is brutal for recommendations: there are few signals per guest, many brand-new sessions with no history at all, and an inventory so large that a generic "most popular" list buries the handful of sites that would actually suit the person looking. The result is a worse experience for the guest and revenue left on the table for the marketplace.

Good personalization here is not about a clever model so much as about doing the right thing under hard constraints - sparsity, cold starts, and a long tail - while never breaking the rules the business depends on. A recommendation that suggests a lower-rated site than the guest last booked, or surfaces an excluded partner, is worse than no recommendation at all.

A pure popularity model is safe but bland; a pure machine-learned model is powerful but opaque and hard to govern. This platform blends both: collaborative-filtering co-occurrence, content/facility similarity, live behaviour signals, review quality, freshness and loyalty, combined into a single re-rank - then it applies hard business rules and a confidence-aware threshold so the picks stay relevant, explainable and compliant.

Those rules are first-class, not an afterthought: recommended sites must meet or beat the guest's last-booked rating, must carry pet-friendliness when it matters, and must respect partner exclusions. Every deterministic stage runs at $0, and every pick carries an honest "Why this?" explanation a guest - or an auditor - can read.

This is a faithful, downscaled reimplementation on the shared synthetic 500-campsite catalog - never the production system. It proves three triggers off one engine: live on-site rails that re-rank as a guest browses, a CRM-style campaign builder over simulated five-year booking history, and a booked-out recovery flow. It deliberately simplifies full-scale email and real-time scoring infrastructure, real booking/payment integrations, model retraining and multi-language depth - all labelled in Architecture → Out of scope. The client is not named.