How Musinsa Built a Self-Optimizing Semantic Layer on Databricks

Musinsa, South Korea's #1 fashion platform ($3.4B GMV, 8,000+ brands, 10M+ users, 13 countries), faced an architecture trilemma where every pipeline forced trade-offs between productivity, latency, and cost. The root cause was not technology but a mental model: "build a table before you analyze." This caused Organizational Amnesia: marts preserved results but lost the logic. We built a TVF-based Semantic Layer on Databricks SQL, shifting from "which table to query" to "what logic to compose." Four time-standardized TVFs, demand-driven caching (CTVF) on Delta Lake, and three AI agents (Genie, MCP) enable self-optimizing analytics managed via Unity Catalog. Results: 75% fewer managed assets (1,000 to 250), 98% faster analysis (7 days to 2 hours), 3.6x self-service increase (15% to 55%).Attendees leave with a trilemma framework and concrete TVF patterns. Includes a live SQL demo.