Agoda built a multimodal content system that unifies over 700 million hotel images and multilingual reviews (40+ languages) through a shared topic taxonomy. Instead of separate pipelines for images and text, both modalities are mapped to common topics like Pool, Breakfast, and Room Quality. Image classification models generate semantic labels while NLP pipelines extract review phrases and sentiment, all pre-aggregated offline into topic-level packages. PySpark jobs orchestrated via Kubeflow handle distributed processing, with Couchbase serving as the low-latency retrieval layer. The offline-first design trades freshness for performance and requires careful governance of topic definitions to prevent semantic drift across languages.
167 Impressions