We solved a fundamental search engineering challenge: running ML models (transformers, neural rankers) and gradient boosting algorithms (LightGBM, Catboost) at native C++ speed.

Modern commerce search demands both rapid ML iteration and millisecond latency at scale. Data scientists want to deploy new transformers and neural rankers daily. Our system serves billions of queries during Black Friday Cyber Monday (BFCM). Traditional approaches force a choice: flexibility or performance. We refused to choose.

Our Search team operates infrastructure serving millions of merchants globally. We handle real-time indexing, retrieval, ranking, and experimentation. Our ranking pipeline combines classical information retrieval with modern ML—typo correction, synonyms, faceting, and semantic search across storefronts and the Shop app.

We built privacy controls, multi-language support, and accessibility features into the core architecture. Merchants get merchandising APIs and relevance controls for customization at scale. Automated metrics and evaluation pipelines let us ship ranking improvements without breaking production.

Our solution: RankFlow, a Domain-Specific Language, and TurboDSL, an execution engine enable data scientists to deploy models trained on billions of queries in minutes. We get the best of both worlds: Machine Learning (ML) innovation and systems performance.