Problem
Rippling's go-to-market (GTM) team of close to 2,000 people had to write SQL to build their account books and ingest data themselves. Vendor data on one person arrives from many sources with duplicates and generic attributes. The old batch stack (orchestrator, dbt, legacy cloud warehouse) had delivery SLAs on the order of 24 hours, and LLM jobs on it were slow and expensive. Connecting a chat tool straight to the warehouse recomputed embeddings per question and lost deterministic filtering.
Example from the article: a marketer wants to know how a Super Bowl ad in January 2026 affected down-funnel metrics, without writing queries.
Solution in brief
Rippling built a Databricks "GTM AI database": incremental ingestion, ML entity resolution into one golden record per person, LLM tags, and a hybrid index (filters plus embeddings) computed once. Agents in a chat app, GrowthOS, read from it. Early A/B tests showed lifts in demos booked and opportunities.
Questions this case answers
- How to deduplicate records of the same person from many vendor feeds without comparing all pairs, using blocking keys and a supervised match model
- How to run LLM classification and summarization over a warehouse incrementally so that only new or changed rows are processed and embeddings are computed once
- How to deduplicate the same person across several data vendors without comparing all pairs
- How to serve hybrid (filter + semantic) search over warehouse data without recomputing embeddings per question
Tools and methods
- Entity resolution (blocking + supervised matching) — dedupe people across vendors into a golden record with a confidence score
- LLM summarization and classification — structured tags from raw person and company data
- Databricks Vector Search (storage-optimized) with Databricks GTE embeddings — hybrid search: predicate filters first, approximate nearest neighbor (ANN) on 1024-dimension embeddings
- Delta Live Tables, change data feed, Iceberg compatibility — incremental processing, cross-engine access
- Genie API, Databricks AI/BI — analytics and campaign analysis
Google or a one-time email code. No password.