Systems & Graphics Engineer specializing in WebGPU compute, browser-native LLM runtimes, and high-performance memory architectures. I build infrastructure that forces the browser to run code it wasn't designed for.

⚡ Selected Open-Source Work

🎮 WebGPU & Compute Shaders

• cr-vector-sync — WebAssembly/WebGPU local-first Vector Database with E2E encrypted peer-to-peer CRDT syncing via OPFS.
• paged-attention-wgsl — High-throughput PagedAttention & block-tiled KV-cache virtual memory engine for running multi-head transformer models on WebGPU.
• flash-attention-wgsl — Hardware-tiled Online Softmax FlashAttention-2 compute engine in WebGPU & WGSL. Eliminates O(N²) intermediate attention matrix allocations with workgroup SRAM cache tiling and causal masking.
• webgpu-vram-pager — Virtual memory paging & ring buffer for executing 8B+ LLMs in-browser, bypassing WebGPU maxStorageBufferBindingSize limits.
• radiance-cascades-wgsl — Real-time flat & nested radiance cascades for global illumination and volumetric light transport.
• subsurface-scattering-wgsl — Multi-layered separable BSSRDF approximations with Gaussian convolutions for human skin rendering in WebGPU.
• ocean-fft-wgsl — Radix-2 Fast Fourier Transform for real-time ocean wave simulation (Phillips spectrum) in WGSL compute shaders.
• volumetric-clouds-wgsl — Multi-scattering raymarching through Worley/Perlin noise density volumes with Beer's Law attenuation.
• marching-cubes-wgsl — GPU-accelerated isosurface extraction and procedural terrain generation with prefix-sum compaction.

⚙️ Data Structures & Low-Level Infrastructure

• splat-bvh-core — Parallel Bounding Volume Hierarchy constructor and frustum culling engine for 3D Gaussian Splatting rendering.
• swarm-refactor — Distributed multi-agent execution framework built on Node worker threads with shared ArrayBuffer state locks.
• hyper-hnsw-core — Browser-native Hierarchical Navigable Small World vector search implementation.
• helix-lsm — Log-Structured Merge tree optimized for WASM runtimes (edge KV emulation).
• nova-wasm — Portable sandboxed execution environment built for zero-trust multi-tenant serverless nodes.

🧰 Middleware & Network Architecture

• edge-rbac-wasm — Ultra-fast stateless Authorization middleware compiled to WASM for Cloudflare/Vercel Edge.
• auto-heal-selectors — Zero-dependency semantic DOM auto-healer proxy for Playwright and Puppeteer.
• agentic-memory-kv — High-speed strictly typed Key-Value store mapped over SharedArrayBuffer for multi-agent LLM systems.
• edge-context-router — Ultra-low latency V8 isolate request router for multi-tenant edge AI API gateways.
• idempotency-middleware — Distributed Redis-backed idempotency lock system to prevent race conditions on multi-node inference endpoints.