What I'm building
Mila runs Gemma 4, Llama 3.x and GPT-2 on the consumer hardware you already own, from explicit neural-network components rather than a framework. In development since 2021, currently in public beta.
Project site
mila.toddt.me
What Mila is, how it works, and the reading path through the
code. Design writeups on quantization, attention and CUDA kernel work.
Read the documentation →
Source
ToddThomson/Mila
C++23 modules and hand-written CUDA. MIT licensed. Models
reproduce their HuggingFace references token-for-token at the target precisions.
Browse the source →
Discussion
Design and progress
Where the engineering decisions get argued out in public —
measurements, dead ends, and what shipped.
Join a discussion →
Earlier work
Before CUDA I spent years on the TypeScript toolchain — compilers, bundlers and minifiers, back when the ecosystem was still deciding what those should look like. These are archived or dormant, kept up for anyone still reading them.
tsprojectArchived
TypeScript minifier and modular bundle optimizer, built as a gulp adapter.
Ts2JsDormant
A small, fast, extendable TypeScript compiler supporting incremental and type-checking builds.
TsMinifierDormant
Minifier providing identifier mangling and whitespace removal.
jQuery Mobile Subpage WidgetArchived
A subpage widget for jQuery Mobile, from an earlier era of the web.
Contact
I'm a single developer, and I read everything that comes in. Design questions about Mila are best raised in its discussions, where the answer stays useful to the next person. Anything else, email works.