Customer Master Data — Strangler Migration

Replaced a mission-critical Customer Master Data monolith with zero downtime via the Strangler Fig pattern: a thin adapter layer (anti-corruption layer) kept exposing the legacy SOAP and REST contracts while delegating to the new platform API, letting each consumer migrate method-by-method at its own pace behind feature flags (instant rollback). Migration validated by high-coverage functional parity tests comparing, endpoint by endpoint, the legacy response vs the flag-on response (golden master) — equivalence proven, not assumed.

Designed and operated end-to-end ETL + data pipelines feeding Snowflake and a data lake layer (raw + curated) for downstream analytics and master-data lineage.

Automated a human-run corporate-client deduplication workflow with LLMs, eval-first: before writing the first prompt, collected 1,000 real human-decided cases as a golden dataset; every prompt candidate ran against it and produced accuracy statistics, and the same harness became the regression gate for model-version migrations. AI accuracy exceeded the measured human baseline — which is what justified automating.