ADITYANPbuilder. operator. data.Book a free teardownBook teardown

Cases · Data lead · Enterprise B2B · anonymized

From a 30-year private-cloud monolith to a reusable cloud data platform

A horizontal, decades-old data platform with manual runs and no quality gates. Redesigned as cloud-native data products — so multi-day batch cycles finished in hours, and the same frameworks fed the website, analytics, and ML.

Target architecture — metadata-driven flow
Runtime · Dagster + MLflow on Kubernetes · AzureSourcesLegacy + feedsMetadata driverOne reusable moduleFrameworksIngest · Quality · StandardizeTransform · Automated QAPySpark · PandasData productsFact · DimensionOwned domainsSame transformed products feedWebsiteAnalyticsML / AIGoverned inputsBefore: private-cloud monolith · manual runs · no QA · weeks of processing

The problem

  • The data platform was a horizontal monolith on a private cloud — layered for roughly 30 years.
  • Processing was mostly manual. No automation. No QA checks in the path.
  • Everything lived in one tangled flow, so teams could not own a slice of the data or ship changes safely.
  • Downstream consumers — the main website, analytical dashboards, and ML/AI workflows — depended on that fragile path.

What I did

  • Redesigned the end-to-end data process to run on cloud, with a clear path from ingestion to publish.
  • Split the estate into logical data products so each product could own its own pipeline and lifecycle.
  • Remodelled the warehouse around fact and dimension design so analytics and products shared one trustworthy shape.
  • Built a metadata-driven platform: one driver module wired to reusable ingestion, quality, standardization, and transformation frameworks.
  • Automated QA checks in the path — not after someone noticed a broken dashboard.
  • Ran processing with PySpark and Pandas, orchestrated through Dagster.
  • Migrated data from the private cloud to Azure with validation and cutover discipline.
  • Stood up Dagster, MLflow, and the processing stack on Kubernetes — platform and pipelines, together.
  • Partnered with AI engineers so model workflows received correct, governed inputs — not ad-hoc extracts.

What changed

  • Data processing that previously took multi-day batch cycles (often spanning a week or more end-to-end) completed in hours after the rebuild.
  • Reusable frameworks replaced one-off scripts — new pipelines plugged into the same driver.
  • Website, analytical dashboards, and ML/AI workflows all consumed the same transformed products.
  • Quality gates and data-product boundaries made ownership and change safer across teams.

Stack

  • Azure
  • PySpark
  • Pandas
  • Dagster
  • MLflow
  • Kubernetes
  • Fact / dimension modelling
  • Metadata-driven pipelines

Need a similar rebuild?

Book a free 30-minute teardown. Bring the current stack. Leave with a clearer next step.