Swiggy's engineering team describes their data-centric pipeline for building and enriching Points of Interest (POIs) to solve the 'last-last-mile' delivery problem. Rather than inferring both boundaries and semantics from raw address clusters (a computationally expensive approach using hierarchical agglomerative clustering and DBSCAN), they now use open building footprint datasets (Google Open Buildings, OpenStreetMap, Microsoft Global ML Building Footprints) as the geometric foundation and enrich them with Swiggy's internal order and address signals. An LLM-based tagging system resolves noisy customer address text, web search context, and reverse-geocoded data into structured POI names and types, achieving ~85% accuracy at ~1,000 POIs per 4 minutes. A separate gate identification pipeline uses historical GPS delivery traces to find real entry points into complexes, reaching ~70% accuracy before human validation. The result is a modular, scalable system that separates geometry discovery from semantic enrichment.
