01 · selected work
A corpus of structural negative results on when machine learning helps the Capacitated VRP, from a research internship at HEC Montréal. Includes a classical Rust engine that beats HGS.
A 6DOF flight sim and autopilot in Rust. A quadrotor and a fixed-wing, flown by complementary/MEKF/INS estimators and PID/LQR control over a 1/1000-scale Earth.
Android runtime that runs LLMs, diffusion, and speech models locally on real phones. One Kotlin API over llama.cpp, stable-diffusion.cpp, whisper.cpp, and ONNX.
Local image-generation system. Pipeline core, queueing server, browser UI, tuned with Xformers, BFloat16, WaveSpeed, and Stable-Fast.
Offline-first Android reader for web chapters, PDF, EPUB, and manga, with on-device chapter summaries through llmedge.
02 · about
I build out of curiosity, or to get rid of a problem that's bugging me, and most of it ends up being about running heavy compute locally.
It started with small stuff, a chess engine in C and the controls for an autonomous submarine, then kept drifting toward local AI. That's the work I care about most: LightDiffusion, a Stable Diffusion stack I tuned to be one of the fastest, and llmedge, an Android runtime that runs LLMs, diffusion, and speech right on the phone. The rest tends to come from everyday friction: Emaki because I wanted a better reader, with the summaries running on-device, and FrameLimiter because I was tired of my Air cooking itself in games. Pilotrs is a flight simulator core I wrote in Rust to play with control loops: keeping a relaxed-stability airframe (which stalls in 0.3 seconds on its own) airborne using noisy sensor estimates rather than clean simulator state.
off the clock · top 0.03% EU in Valorant · Model UN (C-MIMUN)
03 · index