I build inspectable AI agents and local-first tools.
Guarded automation · observable execution · practical software
把复杂工作流做成可运行、可观察、可信任的工具。
Projects · QianPulse · Linux-Agent · MacKit · Caixuntong
🇨🇳 中文版（点击展开）
我做能使用工具但不沦为黑盒的 AI agents：动作应当可见、危险操作必须有明确边界、失败必须可恢复。
我的项目通常落在三个交集上：
- AI agents —— 工具调用、受护栏保护的执行、人类可读的结果
- 本地优先软件 —— 快速、把控制权留在用户手里的实用工具
- 信号与数据工程 —— 从嘈杂的真实世界数据中提取可信结构
QianPulse — 把日常车流变成桥梁监测网络
不新增一个传感器：过桥的车本来就带着 IMU，每次通过都是一次等待被记录的测量。QianPulse 把成百上千次嘈杂的车辆穿越融合成每座桥的频率指纹，筛查持续的响应偏移——告诉养护人员哪座桥值得优先检查。已在缩尺结构（真实 iPhone）与真实车载实测数据上完成验证。
Python Streamlit 信号处理 Welch PSD JS散度 MapLibre AGPL-3.0
用自然语言操作 Linux：受护栏保护的执行、可回滚、结果可观察。
Python LLM agents 自动化 运维
原生 macOS 菜单栏工具箱：剪贴板历史、端口管理、Finder 路径、临时文件传输。
Swift AppKit SwiftUI macOS
市场情报与模拟交易：显式风险边界 + 本地只读控制台。
Python 行情数据 回测 风控
我创立并持续迭代的金融资讯产品。
产品开发 金融资讯 Web
- 让动作可观察，而不是藏在精致界面背后。
- 给有真实后果的操作加上明确护栏。
- 偏好小反馈回路、可逆变更、能跑起来的原型。
- 让系统始终能被理解、调试和改进。
- 区分证据与模拟——每个数字都应可追溯。
- 华南理工大学 计算机科学与技术，辅修金融
- 独立 Apple 开发者，多款 App 已上架 App Store
- 黑客松选手，喜欢把粗糙想法快速做成能用的产品
- WorldQuant Global Challenge Top 100
正在推进 QianPulse（在线演示 + 开源）、构建更安全的工具型 agents、打磨 macOS 本地软件。
I work on AI agents that can use tools without becoming black boxes: their actions should be visible, risky operations should have clear boundaries, and failures should be recoverable.
My projects usually sit at the intersection of:
- AI agents - tool use, guarded execution and human-readable outcomes
- Local-first software - fast utilities that keep users in control
- Signal & data engineering - extracting trustworthy structure from noisy real-world data
QianPulse — turn everyday traffic into a bridge-monitoring network
No new sensors: vehicles crossing a bridge already carry IMUs, and their crossings are measurements waiting to be recorded. QianPulse fuses hundreds of noisy vehicle trips into a per-bridge frequency fingerprint and screens persistent response shifts — telling maintenance crews which bridge deserves inspection first. Validated on a scale model with a real iPhone, and on real drive-by field data.
Python Streamlit signal processing Welch PSD JS divergence MapLibre AGPL-3.0
Natural-language Linux operations with guarded execution, rollback and observable results.
Python LLM agents automation operations
A native macOS menu-bar toolbox for clipboard history, port management, Finder paths and temporary file transfer.
Swift AppKit SwiftUI macOS
Market intelligence and simulated trading with explicit risk boundaries and a local read-only console.
Python market data backtesting risk controls
A financial information product I created and continue to develop.
Product development financial information web
- Make actions observable instead of hiding them behind polished interfaces.
- Put explicit guardrails around operations with real consequences.
- Prefer small feedback loops, reversible changes and working prototypes.
- Keep systems understandable enough to inspect, debug and improve.
- Separate evidence from simulation — every number should be traceable.
- Computer Science and Technology at South China University of Technology, with a minor in Finance
- Independent Apple developer with multiple apps published on the App Store
- Hackathon builder who enjoys turning rough ideas into working products quickly
- WorldQuant Global Challenge Top 100
Shipping QianPulse (live demo + open source), building safer tool-using agents, and improving local software for macOS.