All Projects
Claude Daemon
Running Claude Code as an isolated macOS user with browser access
Visual Acuity Test for LLMs
Evaluating AI models' ability to read fine details in high-resolution images
flex-nano-vllm
A minimal vLLM inference engine built from scratch using PyTorch FlexAttention
Project MIST
Building the MIST robot prototype with AI-assisted design
Agent-Environment Middleware
A plugin architecture for composing LLM agents with middleware patterns
voyager-mcp
Voyager MCP
A Model Context Protocol (MCP) server for Claude Code integration
Claude Code System Prompts
A website that shows Claude Code's system prompts across different versions
wikimcp
WikiMCP
A MCP server to let Claude explore random Wikipedia pages
uvx llmcp serve
LLMCP
A minimal MCP server for LLM to query other LLMs via LiteLLM and MCP.
LLMProc
A Unix-inspired framework for building robust, scalable LLM applications
fork()
Forking an AI Agent
A MVP exploring fork() pattern for AI agents
kodx
Kodx
Docker-based AI coding agent inspired by GitHub Copilot Codex
Santa Hat AI
A webapp that uses MediaPipe face detection to automatically place festive santa hats on profile pictures
vFLUX
An optimized FLUX model inference engine
T5 FlexAttention
T5 model optimized with FlexAttention
AI Shell
A transparent shell wrapper for building context-aware AI tools
Additive Rotary Embedding
A competitive variant of rotary position embedding (RoPE)
Multi-head Latent Attention
I implemented Multi-head Latent Attention from deepseek-v2
Mixture of Depths
I implemented Mixture of Depths from Google DeepMind's paper
minLoRA
A minimal library for LoRA (200 LoC!), supports any model in PyTorch
Flex Diffusion
I fine-tuned Stable Diffusion 2 for dynamic aspect ratio generation
DDIM inversion notebook
My popular notebook demonstrating DDIM inversion using Stable Diffusion
Anim·E
State-of-the-art anime image generator at the time, before Stable Diffusion fine-tuned models