Seven years building petabyte-scale data platforms and the production LLM agents that run on top of them. Most recently Walmart Global Tech, before that AWS. Building Vezilo after hours. Photographer, traveler, nine-ball on Sundays.
Six balls, one cue. Move the mouse to aim, click to shoot: the further from the cue ball you click, the harder it hits. Pot them all, or scroll on. Nobody is keeping score.
Seven years, newest first. Keep scrolling, each role pots a ball, ending on the black.
Two agents shipped on top of the platform: a triage agent that reads on-call tickets, retrieves similar resolutions and runbooks, and safely closes the ~30% that never needed an engineer, and a sprint-report agent leadership reads every week. Before either touched real data I built the RBAC layer that made pointing a model at production safe. The model drafts, the code decides.
Led the deprecation of a legacy customer identifier across petabyte-scale datasets and three partner teams over four quarters, with zero data-quality regressions. Risk-scored sequencing, shadow-mode parity validation at every stage, four engineers on independent tracks, two of them mentored to promotion.
Architected the petabyte-scale table format migration: parallel tables with dual-write, hidden partitioning, consumer-by-consumer cutover with rollback until the last team moved. 64% runtime reduction, and schema evolution unblocked downstream. Alongside it, $100K a month of cloud compute removed, then another 15% with serverless Spark.
Owned the Scala and Spark pipeline that turns billions of search, cart and purchase signals into the datasets three consumer teams run ad targeting, audience segmentation and ML measurement on. Re-architected the critical batch jobs for a 90% runtime cut and moved three high-traffic Ads APIs to async flows for 67% lower p95.
Event-driven Java services on Lambda, Step Functions and DynamoDB processing enterprise billing data, plus a validation library other teams adopted to stop invalid account relationships reaching invoices. Idempotent consumers everywhere: you cannot process money one and a half times.
Two tracks that keep feeding each other: the platform underneath, the AI on top.
My own app: product, mobile and backend, all of it mine. Still very much on the table, and the place I get to be unreasonable about details. More to share once it is ready to be seen.
Work in progress · details soon
Photographs from the road. Keep scrolling and the wall slides past.
Where the rest of the week happens.
In a league now, playing every Sunday. Position play is planning three shots ahead, which is most of the day job too.
Long drives and long weekends across the US, cities in India, and one very good week in Mexico. Camera always in the bag.
Sunnyvale is home base. Coast on Saturday, league on Sunday, and the table is always open for the next frame, wherever it is.
The table is always open. Sunnyvale by default, happy to play anywhere in the US, and always up for a conversation about real engineering at real scale.