The Lab · 10 projects
Projects
What I build when nobody is asking for it.
01
LLM Leaderboard
Benchmark scores, prices, and release dates for today’s language models, pulled from fourteen sources onto one page every night. It answers one question: which model should get the work.
02
Fighting Feel
A beat-synced gun-fu prototype set in a Kowloon tower in 1987, built to show character, camera, and controls. Every press is judged against the beat.
03
Galaxy simulation
A galaxy that runs in a browser tab. Gravity, gas, new stars, and the drawing all stay on the graphics card.
04
Hand Walker
A severed hand that walks on its fingertips, built in Three.js. The walk is solved from the anatomy of the model each frame, and the hand lives in a small room you can play in.
05
Close Combat-like
A WW2 tactical sim at company scale, in the Close Combat lineage. An order is a request: each team reads the ground for itself and can refuse. 247 soldiers a battle.
06
Life-sim
An interactive evolution playground. Neural-net swimmers that learn within their own lifetimes, under purely economic selection, with no fitness function. WebGL2.
07
OceanSim
A native Godot sailing simulation with a reactive ocean, boat wake, volumetric sky, and a beach shaped by receding waves.
Also
08
Atlas
My own assistant: an Android app on my phones, with a Python agent on a VPS behind it that does the work. It builds, tests, and installs its own updates once I approve them.
09
Automated trading system
A trading system that runs on a VPS against a Dockerized Interactive Brokers gateway. Orders go out as brackets. Prices are checked against more than one source, and stale or contradictory data is rejected before anything trades. Responses are cached.
It has only ever traded paper. I test it forward, on live markets with paper money, in preference to backtests on history.
10
NHL analytics
A personal platform, built for me and about nine friends. It loads the public NHL API into MySQL, and a Discord bot on top turns a plain question into stats, trades, and analytics. The bot works in steps: it picks a tool, reads the result, and decides what to ask next, where a simpler bot would make one call to a model.
I started it around 2020 and it is no longer active. The code is private.