Systems engineer working on local LLM inference and the infrastructure around it: inference servers built on
llama.cpp, GPU compute, and shipping the same code to native, Python and WebAssembly targets. Mostly Rust and C++,
with TypeScript on the product side.
Experience
Software Engineer, AI Infrastructure · Hyperspace
Apr 2023 – Feb 2026
Distributed AI compute network: contributed consumer hardware running local LLM inference.
- LLM inference server (Rust, llama.cpp). Built a multi-user inference server: continuous batching so
concurrent requests share one loaded model, KV cache management, and request cancellation. Exposed over an
OpenAI-compatible HTTP API and gRPC, with streaming responses covering tool calls and reasoning output.
- Embeddable inference engine (C++, Rust FFI). Wrote a C API for llama.cpp and Rust bindings over it,
packaging it as a library that runs inference inside the host process rather than as a separate server.
Supports grammars, structured output and tool calls, with cross-platform builds.
- Hardware verification system (Rust, WebGPU, WebAssembly). Designed and built a system that verifies
how much memory a remote machine has actually allocated, using BLAKE3 hash-tree commitments so a check costs a
fraction of the original work. Wrote matching CPU and GPU implementations (WGSL compute shaders) that produce
identical results, delivered as a native library, a Python package (PyO3) and a browser WebAssembly module.
Maintained the service verifying this in production and rewrote it from Python to a Rust gRPC service, with
TLS, health checks, Docker images and CI/CD across environments.
- Visual workflow builder (TypeScript). Built a drag-and-drop editor where users wire LLM, tool and
code nodes into a graph and run it. Node inputs and outputs are generated automatically from each tool's Zod
schema, user-written Python runs sandboxed in Pyodide (WebAssembly), and workflows can be triggered by
incoming webhooks from external services.
- Model providers and agent tooling. Built the model layer for an AI browser: one API over OpenAI,
Anthropic, Google, Mistral and locally-run models, with per-model configuration and tests. Wrote coding-agent
tooling in Rust for applying model-generated diffs to files, web scraping with content extraction, and a
safety layer that classifies shell commands and asks the user before running risky ones.
GPU / Compiler Engineer, Pyccel · Mohammed VI Polytechnic University (al-Khwarizmi)
Nov 2021 – Apr 2023
Open-source Python-to-accelerator compiler. Joined as a 6-month intern (Nov 2020 – Apr 2021).
- Built the CUDA backend for GPU code generation, and implemented core language features including
multidimensional array support and Numba-compatible GPU programming.
- Supervised a team of interns on the GPU integration; helped research groups port scientific Python code to
GPU, and taught programming labs as a teaching assistant.
Technical Skills
- Languages
- Rust · C · C++ · TypeScript · Python
- LLM systems
- llama.cpp internals · continuous batching · KV cache & context management · streaming & tool calling · chat templates & grammars · embeddings
- GPU & interop
- CUDA · WGSL compute shaders · wgpu / WebGPU · C FFI · PyO3 · wasm-bindgen
Education
1337 Coding School (42 Network) · Digital Architect
2018 – 2021
Faculty of Science and Technology, Tangier · two years of university study
2016 – 2018
Baccalaureate in Physics · Complexe Éducatif Lixus, Larache
2016
Languages: Arabic (native) · English (professional, daily working language) · French (professional)