I'm a systems engineer. Rust, C/C++, CUDA.
Three years building inference engines, GPU kernels and production services for a distributed AI compute network. Before that, the CUDA backend of a Python-to-GPU compiler at a research lab.
Everything under work runs client-side, in this tab. No backend. No API key. Your browser is enough.
work
| project | status | what it is |
|---|---|---|
| vram-planner | [live] | what serving an LLM really costs in memory — weights, KV cache, batching, checked against real GPUs |
| tensor-compression | [wip] | the generalized ℒ-product from my published research, compressing video on your GPU |
| coding-agent | [wip] | a small coding agent editing files and running Python in a Pyodide sandbox — no server |
experience
| when | where | what |
|---|---|---|
| 2023–2026 | Hyperspace | Software engineer, AI infrastructure. Wrote a continuous-batching inference scheduler on llama.cpp; embedded llama.cpp in-process behind a hand-written C ABI with safe Rust bindings; designed a hardware-verification protocol with bit-identical CPU and GPU (WGSL) implementations, and ran its production service (Rust, gRPC). Also: a visual workflow builder with a Pyodide sandbox, and the multi-provider model layer of an AI browser. |
| 2021–2023 | Pyccel, UM6P | GPU/compiler engineer on an open-source compiler translating scientific Python to C, Fortran and CUDA. Built the CUDA backend; helped research groups port numerical code to GPU. |
Full detail in the cv.
paper
A. El Hachimi, M. Elalj, K. Jbilou, A. Ratnani. “Generalized ℒ-Product for High Order Tensors and Applications Using GPU Computations.” Mathematical Modeling with Modern Applications (M3A 2024), Springer Proceedings in Mathematics & Statistics, vol. 497, 2025. doi:10.1007/978-3-031-89041-3_6
The tensor-compression demo above is this paper, running in your browser.
contact
mouadelalj@gmail.com · github.com/bauom · cv · Morocco, open to remote & on-site