I scale systems
from 1 to 1

Problem

Amazon Hub needed deep, per-seller delivery-eligibility checks inside the hottest paths in e-commerce: the detail page at 1.6M TPS and checkout at 500K TPS. A single slow call there is a global incident.

Move

Precomputed eligibility, pushed decisions to the edges of the call graph, and designed the checks to degrade gracefully instead of failing loudly.

Result

100% coverage of eligible SKUs for self-ship sellers, with no measurable latency added to checkout. That unlocked an expansion projected at $7.6B in GMS.

Problem

Thousands of VMs on end-of-life Ubuntu, an extended-security bill about to land, and no safe way to move a fleet that big.

Move

Built “Versioned Infrastructure”: reproducible, versioned VM images that turned risky migrations into routine rollouts.

Result

1,000+ VMs migrated from 16.04 to 20.04 with zero downtime, and $250K/year in support costs gone.

Problem

Enterprise customers wanted models bigger than any single deployment could serve, with hard isolation between tenants sharing the same platform.

Move

Led Composition of Experts to run 1T+ parameter composite models, and architected platform-wide multi-tenancy and RBAC underneath it.

Result

An industry-first capability in production, and the isolation guarantees that helped land clients like Accenture and Analog Devices.

Problem

Card payments were failing on opaque, inconsistently handled bank response codes. At gateway scale, every fraction of a percent is real money.

Move

Rebuilt response-code handling across every payment channel (VISA, MasterCard, net-banking) into one coherent decision layer.

Result

+2% card-transaction success rate. At national payment volume, that is millions of recovered transactions.