The founder of Selyron,
shown through the work.

The projects below span AI agents, 3D and real-time graphics, and cross-platform desktop apps.

Let’s talk

Projects

Every project is live or released.

LAAS · WebGPU procedural world

A 4×4 km open world generated live in the browser, without any image, model or audio assets.

Highlights

  • An AI-directed engineering experiment: I wrote the brief, visual bar and hard floors; about 99% of the code was written by Claude over long autonomous sessions
  • three.js WebGPURenderer with TSL materials and raw WGSL compute; strict TypeScript with zero any
  • 131,072 GPU particles, hierarchical wind, and the whole world reproducible from one ?seed= parameter
  • Regression testing with headless WebGPU screenshots, pixel sampling and frame diffs driven by Playwright
  • TypeScript
  • WebGPU
  • WGSL
  • three.js
  • Playwright

OPENVZ Space · 3D building editor

A 3D architecture and BIM editor in the browser: draw walls, place doors and windows, lay out roofs and pipework.

Highlights

  • Parametric components: change a wall's thickness and joined walls re-miter, with openings following along
  • The 3D model and the 2D floor plan are two views of the same data
  • Built-in MCP server lets AI assistants read the scene, create components and adjust parameters in bulk
  • Offline-first: models, materials and decoders are served same-origin with no third-party runtime services; fully localized into Chinese
  • TypeScript
  • React Three Fiber
  • WebGPU
  • Next.js
  • MCP

OPENVZ Agent

A local, always-on desktop AI agent with long-term memory, scheduled tasks and real browser actions.

Highlights

  • Windows / macOS installers built automatically from tagged source, with SHA-256 checksums
  • Electron with contextIsolation; credentials encrypted locally via safeStorage / AES-256-GCM
  • Released continuously up to v2.2.0-rc.11
  • JavaScript
  • Electron
  • Playwright
  • Swift

OPENVZ Leads

A local prospecting agent: finds accounts, writes briefs and drafts outreach, and sends nothing until you approve.

Highlights

  • Human review on by default, with a preview of the real next email in the queue
  • Local bilingual FastAPI dashboard with a daily model-usage cap
  • Released v1.3.0 with a Windows installer
  • Python
  • FastAPI
  • Playwright
  • Docker

OPENVZ AI Workforce

A native macOS app to browse, install and manage 270 AI specialist personas across 18 AI coding tools.

Highlights

  • Persona catalogue localized into Simplified Chinese, with the English original kept alongside
  • Native Rust rendering with byte-level output parity tests
  • Built with Tauri 2: small and fully local
  • TypeScript
  • Rust
  • Tauri 2

OPENVZ Lingo

A desktop app for learning vocabulary in the real context of films, TV series and documents.

Highlights

  • Builds vocabulary lists with context from MKV video, subtitle files or PDF documents
  • Built-in player shows words being learned as on-screen comments, with hover translation and one-click saving
  • Spelling practice, subtitle copying and dictation tests
  • Released v1.0.0 with installers for Windows and Apple Silicon macOS
  • Kotlin
  • Compose Multiplatform
  • Rust
  • FFmpeg

What these projects show

AI agents

Long-running tasks, real browser actions and human approval before anything is sent: AI that acts, with results under control.

Desktop apps

Electron, Tauri and Compose Multiplatform desktop apps, with Windows / macOS installers built automatically.

Graphics and performance

WebGPU, WGSL compute shaders, procedural generation and visual regression testing in headless browsers.

Security and local-first

Credentials encrypted on the device, data kept on the user's own computer, and nothing sent without confirmation.

Which recurring task
should work better?

Let’s talk