I build GPU and ML systems for Apple Silicon and the browser: MLX kernels and model ports, browser-native WebGPU inference, and spatial tools for inspecting generated worlds.

• MLX Core: antialiased linear and cubic nn.Upsample and compiled-kernel correctness for negative-strided inputs.
• MLX Audio: streaming speech generation for VibeVoice and predictor caching for Nemotron speech recognition.
• WebGPU Samples: an external Vision Transformer sample with interactive attention visualization.
• vLLM-MLX: lifecycle-managed model residency, including lazy loading, idle unload, cancellation-safe transitions, and public lifecycle state.
• MLX serving: 15 merged contributions to MLX OpenAI Server and five merged contributions to oMLX, centered on streaming, tool calls, reasoning parsers, cancellation, cache integrity, and API compatibility.

See the full contribution index, selected reviews, and independent deployment evidence.

The Kaminos WebGPU Inference Kit brings substantial models into responsive browser applications. Model applications handle loading, queued inference, controls and cleanup, with model-specific options and lower-level APIs available for custom integrations.

• Image generation: FLUX.2 [klein] 4B runs its text encoder, diffusion transformer and image decoder on the browser's GPU.
• Image to 3D: TRELLIS 2 generates textured meshes with browser-native WebGPU, with a guide to model preparation and asset finishing.
• Skeletal animation: UniMate generates motion for different rigged skeletons from text, with an animated GLB export.
• Live hand reconstruction: WiLoR and MobRecon turn camera frames into hand meshes and landmarks. The demo links to local model setup.
• Relightable photographs: Material Photograph combines MoGe geometry and SuperMat materials with an interactive renderer on one GPU.

Build with a model application, or explore the full model catalog, including background removal, segmentation, depth estimation and Gaussian reconstruction. Some ports live in Kaminos and others in standalone repositories; the catalog links to their current source and integration guides.

• mlx-nf4 is a standalone native Metal NF4 extension for stock MLX, verified from a clean source install on M2 Pro and exercised end to end by double-quantized GPT-2 and the 9.3B mlx-ideogram4 image pipeline.
• cosmos3-mlx is a full-source MLX port of NVIDIA's 16B Cosmos 3 Nano world model for text/image-to-video and joint video/audio generation.
• wilor-mlx rebuilds WiLoR hand-pose estimation end to end in MLX.
• trellis2mlx and pixal3d-mlx bring image-to-3D generation to Apple Silicon without PyTorch.

Kaminos is very much a work in progress: it is the research bench where I explore the ideas that interest me most right now, especially browser-native ML, spatial interfaces, and tools for inspecting and shaping generated worlds. The WebGPU inference kit above is the most mature reusable component to emerge from that work.

Contact: n.lyons53@gmail.com