Skip to content
LTIMindtree · Avery DennisonAug 2021 to Dec 2022Software Engineer - Data

Enterprise Master Data Pipeline

Ten million fragmented customer records reduced to 2.5 million trustworthy ones.

10M

Records consolidated

2.5M

Golden records delivered

4x

Governance metric gain

30%

Pipeline efficiency gain

The problem

Customer data lived in regional systems that disagreed with each other. The same customer existed several times over under different spellings, which made unified CRM reporting and any downstream analytics unreliable.

Approach

An Informatica ETL pipeline consolidates ten million records from the regional systems into a single master flow, normalizing formats before any matching is attempted.

Deduplication and automated validation testing run inside the pipeline, which moved governance metrics four-fold. The improvement came from measuring quality continuously rather than auditing it periodically.

Python, Oracle PL/SQL, and REST enrichment lifted legacy pipeline efficiency by 30%, largely by replacing row-at-a-time work with set-based operations.

The result is 2.5 million golden records published to downstream Oracle CRM/CX systems as the authoritative customer view for AI-driven analytics.

Architecture

publishRegional systems10M recordsInformaticaETLEnrichmentPython + RESTDedupmatch rulesGolden records2.5MOracle CRM/CXdownstream
StorageComputeClient
Validation runs inside the pipeline, so quality is measured continuously rather than audited later.

Decisions and tradeoffs

  • Conservative match thresholds

    Prefer leaving a duplicate over merging two real customers.

    Why
    An incorrect merge is very hard to unwind and corrupts downstream reporting silently. A surviving duplicate is visible and fixable.
    What it cost
    A residual duplicate rate that needs periodic manual review.
  • Validation inside the pipeline

    Automated quality tests on every run.

    Why
    Continuous measurement is what produced the four-fold governance improvement; periodic audits only find problems long after they land.
    What it cost
    Longer runtimes and a test suite that must evolve with the data.