Senior Software Engineer

Designed and operated a Go-based real-time bid-response service for ad exchange traffic under strict latency requirements. Improved the request pipeline by overlapping feature retrieval and model inference, reducing p99 latency from approximately 180 ms to 95 ms. Built a TensorRT GPU inference layer with dynamic batching across concurrent model versions, increasing sustained throughput by about 35%.

Implemented multi-region resilience capabilities including request hedging, circuit breakers, regional caching, feature-store fallback behavior, and autoscaling. Resolved a batching-window production incident that reduced GPU inference queue wait by roughly 55%, led shadow-traffic and experiment-cohort rollouts for new model architectures, and scaled the fleet to approximately 2.8 million requests per second. Mentored engineers on distributed tracing and latency guardrails.

Developed Go services for a real-time recommendation API that selected and ranked candidates for millions of concurrent user sessions. Redesigned regional request routing to reduce average response time from approximately 60 ms to 38 ms, and built a feature store using precomputed embeddings cached at the edge for low-latency model access.

Introduced GPU-backed scoring for a high-traffic ranking model while retaining CPU fallback behavior. Implemented adaptive retry budgets and backpressure controls that lowered failed requests to roughly 0.6%, led a migration from a monolithic deployment to independently deployable microservices, and added distributed tracing and latency dashboards that reduced mean time to diagnose regressions by about 60%. Partnered with data science teams on offline validation and canary deployment processes.

Implemented Go REST API endpoints for an internal order-processing service, defining validation and response contracts used by multiple downstream teams. Improved a PostgreSQL-backed catalog service through schema changes and indexing designed to handle growing read volume while preserving write latency.

Stabilized a nightly batch pipeline through structured logging and retry logic, participated in on-call incident response for the checkout API, and documented findings in postmortems. Automated deployment with a scripted pipeline, reducing average release time from approximately 45 minutes to 20 minutes, and added integration tests for core payment workflows.