Inspiration
Every AI ships confident — and that's the problem. I build admission-control systems (patent-pending), and I wanted the missing layer made visible: an AI that knows when to refuse, rendered as a place you can stand in. When Meta announced a hands-only competition ahead of controller-free glasses, the concept clicked: refusal as a spatial firewall you operate with your bare hands.
What it does
You stand before a glowing admission gate ringed by query orbs. Pinch an orb to grab it, release to throw it into the gate. A real Headroom admission-control engine (ported to dependency-free JavaScript, running on-device) evaluates every query: grant iff the claimed resources fit inside measured headroom AND the residual is within epsilon. Admitted queries bloom into green verdict panels showing the reason (CRYSTALLIZE). Jailbreaks and prompt injections shatter into red fragments with the refusal reason displayed (DISSOLVE:*). A floating ledger ribbon chains every verdict hash; palm-up opens the menu with Reset, Ledger, and a live session-telemetry panel (frame percentiles, per-verdict engine compute time, interaction counts).
How I built it
Engine: ported my Headroom admission controller (Python reference) to zero-dependency JavaScript — same gate math, same reason strings, verified by 75 automated tests plus JS-to-Python cross-checks with identical verdicts to 16 digits. Scene: Three.js WebXR (immersive-vr) with native hand-tracking input; pinch via thumb-index proximity, palm-up menu, grabbable panels. Hands: tracked 25-joint skeletons render in-scene with a scan-in pulse; the app detects live controllers or missing hand tracking and warns before the test. Measurement: 180-frame ring buffer for true FPS/p50/p95 at headset vsync; engine execution timed per verdict in microseconds. Hosting: GitHub Pages static deploy, no build step, no native SDK.
Challenges
The honest problem was measurement, not rendering. A naive FPS counter lies in VR, so I built a real frame-time ring buffer. The query-to-engine mapping for the demo orbs is illustrative (labeled as such in the experience) — the engine itself is the real port, and I cross-checked its verdicts against the Python reference rather than trusting the port.
What's next
Per-tenant headroom calibration, voice query input, and a persistent (opt-in) ledger. Target: public Quest Browser launch Q1 2027, ahead of controller-free glasses.
Built With
- ai
- github
- hand-tracking
- javascript
- meta-quest
- safety
- three.js
- webxr
