Mouad El Alj

Software Engineer · Python, C/C++, Rust, CUDA
Morocco · open to remote mouadelalj@gmail.com mouadelalj.com github.com/bauom

Systems engineer from GPU kernels to distributed services: C/C++ and Rust for performance-critical code, Python for ML, libraries and tooling. Five years across compilers, inference engines, accelerators and language-model agent systems.

Experience

Software Engineer, AI Systems · Early-stage AI startup
Feb 2026 – Present
Building autonomous language-model agents that research and trade prediction markets.
  • Agent research tooling and tool calls. Built the layer letting an agent gather live web evidence before acting: a source registry with per-domain allowlisting and prompt-injection defenses, plus the tool-call interface it reaches them through.
  • Evaluation and dry-run harnesses. Backtest harness and a paper-session runner driving the full agent loop against live markets with no wallet attached, so prompt and model changes could be measured before any capital moved.
Software Engineer, AI Infrastructure · Hyperspace AI
Apr 2023 – Feb 2026
Distributed compute network running machine-learning workloads on contributed consumer hardware.
  • GPU and accelerator programming. WGSL/WebGPU compute shaders with bit-identical CPU implementations, so GPU work can be verified on CPU. Tuned memory throughput and batch sizing across hardware; maintained CUDA and Metal builds.
  • Low-level systems and native integration (C/C++, Rust). Wrote a C API over a large C++ inference library plus safe Rust bindings, embedding it in the host process rather than a separate server. Cross-platform builds with GPU toolchain linking.
  • Real-time scheduling and streaming (Rust). Built the scheduler letting many concurrent requests share one model resident on an accelerator: batching, per-request memory-cache management, cancellation, metrics, results streamed over HTTP and gRPC.
  • Distributed verification (Rust, Python). Designed a memory-residency protocol over hash-tree commitments, so a check costs a fraction of the work; shipped that one core to three runtimes: native, Python via PyO3 wheels built in CI, and WASM/WebGPU. Ran the service in Python (gRPC, multiprocessing) before taking it to Rust/gRPC: bounded worker pools, TLS, health checks, Docker.
  • Internal tooling (Python, TypeScript, React). A Python CLI covering node daemon, inference, model and state operations, plus a browser editor for composing and running user-authored processing graphs.
GPU / Compiler Engineer, Pyccel · Mohammed VI Polytechnic University (al-Khwarizmi)
Nov 2021 – Apr 2023
Open-source, pip-installable Python package compiling scientific Python to C, Fortran and CUDA. Intern from Nov 2020.
  • Built the CUDA backend, generating GPU kernels and memory management from annotated Python, plus core language features including NumPy array semantics, multidimensional arrays and Numba-compatible GPU programming.
  • Ported research groups' numerical code to GPU; supervised interns on the backend and taught programming labs.

Selected project

coding-agent · in-browser agentic coding harness
mouadelalj.com/coding-agent
Runs entirely client-side: no server, no API key, nothing leaves the tab.
  • Local inference (WebGPU). Runs a 0.4–8B model on the visitor’s own GPU, over a 43-model catalog with VRAM-aware selection.
  • Structured tool calls, real verification. Every turn is forced through a JSON schema by grammar-constrained decoding; edits apply by content-anchored search/replace through a four-tier match cascade, and pytest runs for real in a Pyodide sandbox.

Skills

Languages
Python (NumPy, pytest, packaging, PyO3, Pyodide) · C · C++ · Rust · CUDA · TypeScript
ML
PyTorch · Transformers · model serving · batching and KV-cache · LLM agents, tool calls, eval harnesses
GPU
CUDA · WGSL compute shaders · WebGPU · GPU code generation · mixed precision · CPU/GPU parity

Publication

Generalized L-Product for High Order Tensors and Applications Using GPU Computations
A. El Hachimi, M. Elalj, K. Jbilou, A. Ratnani · M3A 2024, Springer PROMS vol. 497 · 2025
doi.org/10.1007/978-3-031-89041-3_6

Education

1337 Coding School (42 Network) · Digital Architect2018–2021
FST Tangier · two years of university study2016–2018
Baccalaureate in Physics · Larache2016
Arabic (native) · English (professional, daily working language) · French (professional)