Building in Public
Open Source
Alongside academic research, I build tools at the intersection of AI agent infrastructure, developer safety, and 3D vision. These projects reflect a philosophy: AI-assisted development should be observable, secure, and joyful.
Agent Ecosystem
Developer Tools for the AI Age
A growing suite of tools designed to make working with Claude Code and AI agents safer, more observable, and more powerful. From security scanning to personality portraits.
VibeGuard
Real-time security scanner for AI-assisted coding. Blocks prompt injection, detects secret leaks, and guards against unsafe file operations — integrated as an MCP server or shell hook.
agentop
Lightweight observability toolkit for AI agent workflows. Monitor token usage, trace multi-step execution, and debug agent behavior in real time with minimal instrumentation overhead.
claude-code-harness
Programmatic harness for Claude Code CLI. Enables automated workflows, batch processing, and CI/CD integration without manual interaction.
claude-code-philosophy
Curated collection of philosophy and best practices for working effectively with Claude Code. Opinionated prompting patterns and workflow principles.
PaperAgent
AI-powered academic paper discovery and summarization agent. Autonomously finds, reads, and synthesizes relevant research from arXiv and beyond.
QBot
Intelligent QQ chatbot with multi-modal capabilities. Supports text, image understanding, and an extensible plugin architecture.
Research Software
3D Vision Projects
Research code and tools from my work on object pose estimation, 3D scene understanding, and spatial AI. Some accompany published papers; others are standalone experiments.
BoxDreamer
Generalizable 6-DoF object pose estimation by dreaming 3D bounding box corners directly from RGB images — no CAD models, no category-specific training. Achieves strong generalization across unseen object categories.
CityLayout
City-scale layout generation and GIS file conversion toolkit for urban scene understanding and large-scale 3D reconstruction workflows.
central-voting-ppf
Point Pair Feature based pose estimation with explicit center voting. Improves accuracy through principled center prediction for dense point clouds.