Festus Owumi is an engineering leader and ML systems engineer with experience building production distributed systems, cloud-native infrastructure, Kubernetes platforms, AI/ML serving systems, and GPU inference infrastructure. His work focuses on LLM inference systems, spanning research, contract engineering, and open-source contributions across vLLM and NVIDIA Dynamo. He is an external contributor to NVIDIA Dynamo, NVIDIA’s distributed inference framework for data-center-scale serving, where he contributed a merged Qwen3-32B-FP8 vLLM recipe for disaggregated inference on NVIDIA A100 GPUs, separating prefill and decode across different GPU workers (2x prefill workers with TP=2 + 1x decode worker with TP=4). He has also contributed to vLLM, the widely adopted high-throughput, memory-efficient LLM inference and serving engine originally developed at UC Berkeley. Outside engineering, Festus writes frequently on Medium, where he breaks down complex systems concepts from first principles. When he is not thinking about distributed systems or LLM inference, he can often be found chasing his next running PB.