I'm a Master's student in Computer Science at EPFL, focused on machine learning systems and AI infrastructure.
I previously worked at Oracle Labs Zurich, where I contributed to the open-source projects Wayflow and Agent-Spec. More recently, I worked on the SwissAI Serving Platform at the EPFL AI Center, focusing on reducing LLM serving cold-start times.
Performance & Systems
Data & ML
Tools & Technologies
- Reduced LLM weight-loading time from minutes to seconds on CSCS clusters
- Built integrations with Servekit and SGLang for distributed model serving
- Achieved up to 51× faster weight loading and 11× faster cold starts
- 🔗 https://youssefboughizane.blog/fast-weight-loading.html
- Designed a shared rollout architecture for concurrent RL training jobs
- Integrated multi-LoRA serving with vLLM and verl
- Improved training throughput by 23–56%
- 🔗 https://youssefboughizane.blog/efficient-rl-finetuning-with-multi-tenancy.html
- Built optimized GEMM kernels on NVIDIA T4
- Hierarchical tiling + vectorized memory access
- Reached ~82% of cuBLAS performance
- 🔗 Code
- Ranked #1 on LeetGPU, 5.3× faster than baseline
- Used shared/constant memory + register blocking
- Profiled with Nsight Compute
- 🔗 Code
- Built OCR system for chess scores (CNN-RNN, PyTorch)
- Reduced WER from 44% → 29%
- 🔗 Code