At Stylumia Intelligence Technology, I owned reliability and capacity for a production GPU inference fleet running Knative and TorchServe on GKE. I codified six failure causes into a reusable pre-deploy checklist, and a GPU packing formula held 344/344 requests at 237ms p50 with zero OOM, cutting worst-case endpoint recovery from hours to seconds.
I also built the AI and agent platform used by the engineering organization, including browser-automation and Rust code-intelligence MCP servers and self-hosted OAuth-2.1 infrastructure. Converting per-session stdio servers into shared HTTP daemons cut idle AI-tooling RAM by about 1.8GB.
During a vendor outage, I ran incident command and restored crawl capacity with a static-pool fallback, keeping four enterprise deliveries whole with zero data gap. I also traced cross-continent network degradation to about 60% packet loss on an upstream provider's peering hop.
Earlier at Stylumia, I built layered cloud security detection and moved self-hosted GitLab across a major version and from GKE to OVH with zero data loss. I also built crawling and data-collection services for 100+ retailers; at Paxcom India, I built scrapers and monitoring for 50M+ daily commerce records and optimized PostgreSQL query paths from 8s to 200ms.
