Ilyas Mallah

Software engineer in the Netherlands

I work mostly in Luau, Rust, C++ and CUDA. My main interests are compilers and runtimes, and making code fast, from CPU hot paths to GPU kernels.

hello@ilyasm.dev   Open to internships and contract work.

Selected work

• llama.cpp: up to twice as fast long-context generation on RTX 50 GPUs in review

  Found why attention ran at half speed with a compressed (q4_0) cache on Blackwell GPUs in llama.cpp's default CUDA builds, and fixed it in a few lines.

• Luau compiler and runtime open source

  Contributions to the Luau compiler, its native code generator for x64 and ARM64, and its runtime, under the handle green-real.

• Rolantir co-founded

  Analytics for the Roblox platform: tracks player counts for about 443,000 games. Built the data service and moved the live database to TimescaleDB, taking dashboard charts from 400 to 900 ms down to 3 to 8 ms.

• Dig It co-owned, sold

  A Roblox game with about 70 million visits and a peak of 27,400 players online at the same time.

In progress

• Ulana release in progress

  A compiler that turns WebAssembly into Luau, so programs written in C, Rust or Go, such as SQLite, QuickJS and zstd, run as plain Luau.

• Ganglion in development

  A Rust and CUDA engine that runs whole brain wiring maps as live simulations. It matches a validated scientific reference bit for bit, runs a 165,000-neuron nervous system faster than real time on one consumer GPU, and drives a physics model of the body.

• infdoc in development

  A command-line check that finds silent slowdowns in self-hosted LLM serving. Every rule comes from a measured case.