A real-time customer view across 500+ tables in a regulated mortgage servicer

The Customer Data Platform is a real-time, 360-degree view of a mortgage servicer’s customers, built for call center, operations and fraud teams at Rocket India between 2021 and 2025. I directed a team of 5 engineers and 2 QA to unify 500+ tables from 50+ source systems, on Kafka, Snowflake and Databricks, into a single platform serving 1M+ API calls at peak hour. The work was recognised with the company’s Champion for Our Customers award in both 2023 and 2024.

A call center agent, a fraud analyst and an operations dashboard were each looking at a different, slightly stale version of the same customer, assembled by whichever team had last built a report against whichever source system they had access to. None of those views agreed with each other reliably, and none of them were fast enough to use while a customer was on the phone.

• This was a regulated mortgage servicer, so every source system had its own change windows, its own owning team and its own tolerance for being touched.
• Consolidating 50+ of them into one platform without a freeze window on any of them meant the migration itself could not be the thing that broke production.
• Once the platform served fraud and operations teams, its numbers had to be provably correct, not just fast.

We built the unification on Kafka and Debezium change data capture, streaming updates out of each source system as they happened rather than batching them overnight, and served the real-time view through ksqlDB and a Kubernetes-hosted API layer that now handles 1M+ calls at peak hour.

In parallel, I organised the analytics side around a Medallion architecture, bronze, silver, gold, on Snowflake. The gold layer specifically, pre-aggregated and clustered rather than queried raw, is what cut a common query from about three seconds to roughly 500 milliseconds for 15+ teams who had been working around the old latency for years.

When Rocket Companies acquired the business, I extended the platform to the new parent company within weeks rather than rebuilding it, and separately led the migration of the underlying cloud footprint from Azure to GCP. PI and NPI data is encrypted across both MongoDB and Confluent throughout, which was non-negotiable given what the platform holds.

A real-time customer view across 500+ tables and 50+ source systems, serving 1M+ API calls at peak hour with no freeze window required on any source system, plus a common downstream query cut from about three seconds to roughly 500 milliseconds for 15+ teams who depended on the old Snowflake layer. The team’s work was recognised as Champion for Our Customers in 2023 and 2024.

I would have invested in schema governance before onboarding the fourth source instead of after the eighth. We spent months retrofitting contracts onto producers who had already shipped, and every one of those conversations was harder than it would have been on day one. The technical debt was not in the pipeline; it was in the agreements we had not made yet.