Architecture

ACSI · Hospitality · 2020–Present · Public

PublicRepresentative · synthetic data
Live diagram — signals blend with CF + facility similarity, then rules, threshold and diversity narrow to the top-N slots ($0).

The deterministic spine

Signals → candidate generation → a hybrid blend (collaborative-filtering co-occurrence + facility Jaccard + behaviour affinity + review quality + freshness + loyalty) → business-rule filter → min-score → re-rank → per-region diversity → top-N. Every stage is deterministic and runs at $0 in both cloud and OSS modes.

The one metered stage

The only optional, metered call is the natural-language explanation ('How these are picked'): cloud Claude Haiku or a recorded local OSS model (Qwen3-8B on M4 hardware), behind a fail-closed budget cap that degrades to $0 OSS. The ranking itself never depends on it — the deterministic 'Why this?' chips carry the specifics.

Architecture · Personalization & Recommendation Platform · Abhishek Saxena