Endeavours
some endeavours.
Side Quests
papers and other detours.
Shipped
out in the open.
Manifold
· April 2026Git desktop client. Your codebase is a manifold.
Change one file and it shows you everything quietly bound to it, heat spreading outward through the file graph. The curvature of that graph even marks the bridges your architecture leans on.
Flutter, Dart, and a bit of spectral geometry.
Whisper
· February 2026Encrypted communication over Möbius geometry, entirely in the browser.
Tuck a message inside an ordinary file, or open a peer-to-peer channel through a 16D bond codec that hands both sides the same secret.
Mathematically private.
What Do You Mean?
· April 2026In-game agentic debugger for Minecraft, split between two agents.
The Doctor reaches into a running mod and rewrites its bytecode on the fly, setting soft breakpoints through a Condition Compiler wherever you need them. The Watchdog never touches a thing; it just keeps a quiet eye on the server's tick rate, waiting for the moment something slips.
A college capstone on Fabric 1.21, built test-first.
Minecraft Server Maintainer
· January 2026A 67KB jar that lives next to your server folder. Double-click it and walk away.
It keeps Minecraft, mods, plugins, and datapacks current, and after every update it checks the server still boots, rolling back anything that doesn't. Crash, and it brings itself quietly back, rate-limited so it never thrashes.
No Docker, no web panel, no subscription.
Morithon
· 2024Unreal Engine 5 death-run. 45 students, one semester.
Can you beat your friends?
Artifacts
built to the edge of an idea, then set down till the tea comes.
two sides of the same coin.
wick is a corpusless corpus; séance reads a model's raw weights like one. two takes on the same heresy: cold weights, cold corpus; hot math, hot attention. and they raised each other.
wick
point it at a folder, ask, and back comes a focused packet of your own passages, each tagged with why it's there. one SQLite file, no embedding server, no GPU, no model download. it trains the semantic space on your corpus and runs heat-kernel diffusion over the document graph (math from the 1800s), so it reads structure flat embeddings miss, and it'll even tell you what's missing. ~1ms warm.
séance
run any LLM straight off cold SSD. no GPU, no framework, zero weights in RAM. a 197KB rust binary memory-maps the weights and reads them one matmul at a time, so a 9B model (19GB on disk) runs in under 2MB of RAM, across 12 models and 10 architecture families. it's IO-bound today (~0.5–14 tok/s), but speed scales linearly with drive bandwidth: no diminishing returns, no cliff. a RAM runtime goes “fast → doesn't run” the moment a model outgrows memory; séance goes “fast → a bit slower.” push compute closer to storage and it laps the traditional runtimes.
séance turns weights into a readable corpus, which is exactly the thing wick pretends to have. each one's question was the other's answer.
Behind the scenes
working tools, not yet released.