Migrating an Oracle finance warehouse to Snowflake in 6 weeks

The team owned a 12-year-old Oracle finance warehouse — five core entities (customers, accounts, orders, invoices, payments), thousands of downstream consumers, and a brittle nightly Informatica job that everyone was afraid to touch.

The plan, described in one paragraph

The director of data opened the AI Migration Planner and described the migration in one paragraph. Within a minute the planner had drafted a structured plan: source connection, portable ontology, mapping scope, target codegen, orchestrator choice, and an 8-point readiness checklist.

The portable ontology

Customer owns Account · Account places Order · Order invoiced as Invoice · Invoice paid by Payment. The ontology was target-neutral on purpose — when, six months later, the team wanted to evaluate Databricks, they re-ran codegen against the same ontology. No rewriting consumers.

Auto-mapping with real transforms

Fifteen mappings auto-generated with explicit transforms — CAST(CUST_ID AS NUMBER), TRIM(FULL_NAME), COALESCE(EMAIL,'unknown@finprd.local'), env-conditional SHA2(SSN) for non-prod, and SCD2 effective dates on Customer and Account. The team bulk-approved high-confidence rows; reviewed only the four flagged for nulls or dupes.

Reconciliation as a contract

After every load, LucidOntos reconciled row counts, checksums, and key aggregates between Oracle and Snowflake. When a Gold mart failed because currency_code was null on 412 invoices, AI Health Analysis named the column, the date range, and proposed the COALESCE fix — in clear, human-readable terms.

We didn't ship a migration. We shipped a reviewable, reversible system that happens to also have moved our finance warehouse to Snowflake.