Paul De Sainte Agathe

• • Building and deploying a multi-node GPU cluster (12× RTX 4090, 2× AMD EPYC 7702) from bare metal (PCIe topology, BIOS, NUMA). Installing and configuring Linux, NVIDIA drivers, CUDA on large multi-GPU system.

• • Optimizing LLM inference with vLLM (batching, tensor parallelism, latency/throughput analysis).

• • Developed an LLM-based pipeline for detecting software vulnerabilities in compiled C/C++ binaries, evaluated on 30,000 files.

• • Built a Big Data pipeline (Kafka, ClickHouse, etc.) to process logs from real-time streaming solutions (WebRTC and Media over QUIC), contributing to an 8× latency reduction compared to DASH/HLS.

• • Implemented a stats interceptor in Golang for real-time streaming log generation (1 log/second/stream).