Omar El-Sayed

I specialize in taking PyTorch models from training graphs to low-latency inference: torch.compile, ONNX/TensorRT paths, and GPU memory tuning for recommender and ranking towers.

• Cut p99 latency 40% on ranking tower after torch.compile + custom CUDA extension audit
• Owned canary rollout for quantized student models beside full-precision teachers

1. Principal Software Engineer, ML Serving

   Gulf Meridian Commerce · 2019 — Present

   Principal software engineer, ML serving: torch.compile and custom CUDA audits cut p99 latency ~40% on a ranking tower; shadow traffic vs ONNXRuntime with automated logit/KPI diffs. Canary rollouts for quantized student models beside full-precision teachers; rollback playbooks and SLO dashboards. GPU memory tuning for large embedding tables; batching strategies for peak retail traffic.