Building reliable GPU infrastructure.

Hi, I'm Ezgi, an AI Infrastructure Engineer focused on GPU fleet reliability, inference performance, and systems profiling and observability.

ezgi@mig-7g.141gb:~ 
$ whoami
Ezgi Tastan
# AI Infrastructure Engineer

$ cat ~/.config/interests.toml
gpu_infra     = "MIG Slicing, GPU Operator, DCGM, NCCL"
orchestration = "Kubernetes, OpenShift, Helm"
ai_ml         = "vLLM, KServe, LLM Inference"
cloud         = "AWS, GCP, Terraform, Terragrunt"
observability = "Datadog, Prometheus"
security      = "SOC2, Trivy, SecComp, RBAC"
Presentations ::~/gpu

Latest Posts

My writings about GPU fleet reliability, inference performance, and systems profiling.

Observability

Profiling MiniMax-H3 on GB300 with SGLang and eBPF

I used eBPF to profile MiniMax-H3 with SGLang on a GB300 tray, measuring stage timing, Ulysses scaling, concurrency, Grace–GPU fabric traffic, and energy.
Observability

Profiling Voice AI on H100 with eBPF

I traced STT and TTS containers on an H100 with bpftime. The GPU was at 0% while the CPU was doing 169K malloc/s. Here's everything I found.