Entity resolution · Databricks · AI adjudication
Salesforce, HubSpot, and event leads land in Databricks as the same people, spelled a dozen different ways. Deterministic rules resolve the obvious duplicates cheaply; Grove brings an LLM to bear only on the ambiguous matches a rule engine can't settle — and writes nothing it can't defend. This page runs against live data.
Three sources feed the same lakehouse table. A single buyer shows up as
ravi.costa@meridiansystems.com in Salesforce, ravi.costa@icloud.com from an
event scan, and again under Meridian Systems Corp vs meridian systems in HubSpot —
sometimes with a nickname, a transposed letter, or no email at all. Exact-match SQL catches a fraction.
The rest is judgment.
Six stages on Databricks. The deterministic layer does the cheap, high-confidence work; Grove is invoked only for the low-confidence band, then a graph pass resolves clusters. Grove also serves this page — every number below is read from Databricks through a Grove workflow.
Email, phone, company and nickname canonicalization in SQL.
Generate candidate pairs by shared key — not 2.9M comparisons.
Rule features bucket pairs: auto-match, auto-reject, or review.
An LLM judges each review-band pair: same person, with a reason.
Connected components over confirmed matches yield golden records.
Humans confirm or flag merges; corrections feed the next run.
Six leads arrive from a conference scan. Run a dry-run ingest and watch Grove orchestrate the whole flow as one workflow: block the new leads against the existing lakehouse, auto-match the certain ones, and adjudicate the ambiguous ones with an LLM — live. It writes nothing, so run it as often as you like.
| Incoming lead | Company | Source | Outcome |
|---|
Scored against held-out ground truth. Deterministic rules alone are precise but miss most duplicates. Adding Grove on the review band roughly doubles recall while holding precision.
| Precision | Recall | F1 | |
|---|---|---|---|
| Rules only | — | — | — |
| Rules + Grove | — | — | — |
The queue Grove hands a human: each pair it adjudicated, with its confidence and reasoning. Confirm a merge or flag a mistake — flags are written back to the lakehouse and become hard rules on the next run.
Loading the review queue through Grove…
Grove runs the non-deterministic work dbt can't — classification, extraction, entity resolution — beside your existing SQL, with RBAC, per-tenant scoping, and append-only audit. This case study is one workflow; the platform is the same one you'd deploy.
Grove