We were burning through Gong's API to feed AI. Modeling the transcripts in the warehouse with dbt cut token costs 20x.