Data platforms built for the petabyte era.
Kevadia designs and builds the data backbone — streaming ingestion, lakehouse storage, modeled warehouses, and the ML infrastructure on top. The same patterns we operated at petabyte scale at Meta and Amazon, sized honestly for where your company is now.
FIG. 03 — Data PlatformsData systems fail slowly, then suddenly.
Pipelines accrete one urgent hack at a time until every dashboard disagrees, every query is expensive, and nobody trusts the numbers enough to act on them. By the time the pain is obvious, the architecture is load-bearing and terrifying to touch.
THE OPERATING SEQUENCE.
- Phase 01
System audit
We trace lineage from source to dashboard, identify the actual sources of truth, and measure cost, latency, and failure rates. You get a map of the system you really have.
- Phase 02
Target architecture
A blueprint with explicit data contracts, storage and compute decisions justified in writing, and a design that scales one order of magnitude past your projection — not ten.
- Phase 03
Incremental migration
Strangler-pattern cutover with parallel runs and reconciliation checks. No big-bang migrations; your reporting never goes dark.
- Phase 04
Operate and hand off
Observability, alerting, runbooks, and working sessions with your engineers — so the system's operators understand it as well as its builders.
WHAT WE BUILD.
Spark, Kafka, Snowflake, Postgres, dbt, Airflow, and Iceberg or Delta — on AWS or GCP, provisioned with Terraform.
FREQUENTLY ASKED.
What does a data platform engagement include?
A typical engagement covers an audit of your current pipelines, a target architecture with explicit data contracts, hands-on build of the new platform, and a staged migration with parallel runs. You end with infrastructure your own engineers can operate — plus runbooks and observability, not a slide deck.
Do we need petabyte scale to justify this?
No. Most clients come to us at terabyte scale with petabyte ambitions. The value of high-scale patterns is that they fail gracefully and grow without rewrites. We size the architecture to your next order of magnitude, not to Meta's.
Can you migrate our data platform without downtime?
Yes. We migrate incrementally using the strangler pattern: new pipelines run in parallel with old ones, outputs are reconciled automatically, and traffic cuts over source by source once the numbers match. Dashboards and downstream consumers keep working throughout.
What stack do you recommend?
It depends on your workloads, but our defaults are Kafka for streaming, Spark for heavy transformation, Snowflake or a lakehouse for analytics, Postgres for serving, and dbt for modeling. We recommend boring, proven technology and spend the innovation budget on your product.
Our cloud data bill keeps growing. Can architecture fix that?
Usually, yes. Most runaway data spend traces to a few patterns: unpartitioned scans, duplicate pipelines, oversized always-on compute, and storage formats that fight the query engine. An architecture pass typically finds savings in the first audit week, and the redesign locks them in structurally.