Computer engineer building local AI tools, native model runtimes, and Linux software.
I work where model architecture meets quantization, kernels, and memory limits.
Selected projects · More to explore · All repositories
|
01 / MODEL COMPRESSION Convert diffusion checkpoints and text encoders into native ComfyUI formats. Architecture-aware tensor selection, bounded conversion memory, optional MSE optimization, and output validation.
|
|
02 / LINUX SYSTEMS A native Linux boot-media writer with a desktop app and CLI. Reviewed write plans, explicit device confirmation, a separate privilege helper, and full source-byte read-back.
|
|
03 / IMAGE UNDERSTANDING Turn a reference image into Anima prompts locally. PixAI tagging and native MiniCPM INT8 inference produce positive and negative prompts, structured analysis, and tag probabilities.
|
|
04 / SPEECH INFERENCE Voice design, voice cloning, and named-speaker generation for the Qwen3-TTS 1.7B family. Packed W8A8 ConvRot inference, static KV caches, and CUDA graph support with runtime diagnostics.
|
- Stripe homepage reconstruction · An unofficial visual study using Astro, GSAP, and a Three.js shader hero. Includes responsive screenshot comparisons and documented limits.
- Ming Image Prompt Builder · Draw regions, place exact text, and export structured Ming prompts inside ComfyUI.
- Qwen-Image 2.1 Turbo LoRA Tools · Experimental local adaptation of base-model LoRAs for Turbo using a reusable transfer profile.
Earlier quantization work & an upstream contribution
- ComfyUI Mixed Quantizer is my earlier single-file converter, with stable W4A8 conversion and experimental mixed-precision planning.
- OpenCode 2 support for Ask is an open pull request for compatibility with the terminal AI client.
I like following an inference problem through the whole stack: tensor layout, kernel dispatch, attention, memory residency, and the runtime that ties them together. I care about reproducible conversions, useful diagnostics, and benchmarks that explain their limits.
Tools I work with: Python, Rust, C/C++, PyTorch, CUDA, ComfyUI, llama.cpp, and Linux.
Away from code: philosophy, mathematics, anime and manga, storytelling, penguins, and cats.
If a model technically shouldn't fit, that usually makes it more interesting.