Try it · public alpha · webcam
Signal Refinery / HandPiano ↗
Hands-free piano using browser hand tracking, a kinematic guard, dwell, and hitboxes. A desktop browser and webcam are recommended. The demo labels its tuning work in progress.
Portfolio / public index
What I have built and explored
I work across movement, interaction, and reliable software. This page separates things you can try now from the larger systems I am still developing.
01 / Spatial interaction
Turning movement into input
My Spatial Input work connects camera observations to browser actions through normalization, filtering, intent, and a stable event interface. The ambition is to let different kinds of input drive an experience without rebuilding the experience for every sensor. The complete multi-input system remains an engineering direction; the demos below show narrower pieces of it.
Prototype family · not a public release
Pointer, games, and music controls
Other experiments include a gesture-to-pointer bridge, table-tennis interaction, and hand-to-audio controls. I use them to test primitives such as tracking stability, dwell, impact, and continuous control. Public artifacts for these prototypes are still being selected.
02 / Playable browser work
Small experiences, direct feedback
Try it · live browser hub
Omega Games ↗
A collection of small, tap-to-play HTML5 games. These use regular touch and mouse input; they are a separate part of my work from the webcam demo.
03 / AI-agent reliability
Checking the claim of “done”
As I built software with AI agents, I saw recurring errors: work reported as complete before a test ran, drafts described as sent, and self-review that missed the same failure as the original work. This became a separate line of tools and writing about independently checkable outcomes.
Inspect the source · open tools
AgentReleaseGate OSS ↗
An open policy and examples for checking AI-agent claims. The repository includes a claim gate, a reward-hack mirror, and a free field guide.
Read · companion site
Worldweaver ↗
The AI/swarm-facing side of my work: cards, public context, dated world-state views, and regeneration paths. Its observations have their own dates and limits.
04 / Independent research
Reproducible work, stated limits
Read and verify · September 2026
P-384 scalar inversion candidate →
A 422-operation addition chain, frozen certificate, multiple reconstructors, and a written account of the search. This is a candidate under a defined operation model; it is not a claim of runtime speed or global optimality.
Public index updated . Project stages and links can change; a demo is evidence of its own behavior, not every proposed use of the larger system.