Jerry Lee

Wuhan University · Computer Science

Jerry Lee

I build small, dependable tools across systems, AI agents, computer vision, and the web.

Explore work GitHub
Based in Wuhan, China
Working across Systems · AI · Vision
Current direction Useful tools, clear evidence

01 / Selected work

Built to be inspected.

Projects that favor explicit behavior, reproducible workflows, and code that can be understood without guesswork.

ipureread-only · offline-capable
$ npx ipure sub ./nodes.yaml --probe --html report.html

   节点 5 · 独立出口 IP 5 · 实测可用 5/5
   纯净度 ▇▇▇▇▇▇░░░░ 61/100
   1  US-NY-01  http  158.46.131.10  家宽  原生  100 S
   2  TW-TP-07  http  178.171.80.225  机房  广播   45 C

Network diagnostics · Node.js

ipure

Local-first IP purity check: itemised 0–100 scoring, real-traffic exit verification, duplicate-exit detection, native/broadcast via RIR data, DNSBL and Spamhaus DROP. Live demo report.

  • Zero dependencies
  • Exit verification
  • Spamhaus DROP
agentjanitorread-only scan
$ npx agentjanitor

   检测到 4 个 AI CLI · 数据共 1.9 GB
   🚨 3 个密钥泄露进会话记录 (transcript)
     DeepSeek  sk-…a1f2  ×5 处
   ⚠ 444 MB 超过 90 天未动 — 可清理候选

Security audit · Node.js

agentjanitor

The health check for every AI CLI on your machine: disk hogs, plaintext keys in configs, secrets that leaked into session transcripts, and MCP config duplication across CLIs.

  • Secret scanning
  • 14 CLI registry
  • Read-only
dueldeterministic arena
$ duel eval bots/counter.mjs bots/adapt.mjs --games 100

   counter   86.0%   adapt   8.0%   draws  12.0%
   avg 6.5 turns/game · 0 invalid moves

AI agents · Node.js

duel

A deterministic AI-agent arena in your terminal: any executable can fight, every match is a replayable 2 KB file, and the ELO ladder keeps receipts.

  • Replayable
  • ELO ladder
  • Pick&ban
tapeoffline · no keys
$ tape gif demo-session.jsonl --delay 60

🎞 25 frames · 800x313 · 233 KB (delta)
   every frame is a step — replay, grep, share

AI agents · Node.js

tape

A time-travel debugger for AI agent sessions: reads the JSONL your agent already writes, replays any step as frames, greps across formats.

  • Time travel
  • Zero deps
  • Delta GIF
mogwaia pet with a repo for a brain
   /\_/\
  ( ^ω^ )   ← Gizmo, feeling cozy
   > ~ <
   growth rings: 5 commits · bond: 5/100
   curiosity  ▇▇▇▇▇▇░░░░ 55%

AI agents · Node.js

mogwai

A terminal pet whose brain is a git repo: memories decay, personality drifts with what actually happens, and forgetting is reversible.

  • Local-first
  • Memory decay
  • Tamagotchi
synapseprobe → verdict
$ synapse doctor --fix

   station-1  ok    214ms  12 models
   station-2  fail  ECONNRESET (3h)
   预览: chmod 664 → 600  ×38
   (--apply 真改，自动备份)

Network diagnostics · Python

synapse

Connectivity monitor and config repair for model relay stations — pure Python stdlib, probe-first, and it never edits a config without a backup.

  • Stdlib only
  • Probe-first
  • Auto backup
agentic-risk-scanscan complete
$ agentic-risk-scan scan .

risk score     100 / 100
findings       1 critical · 11 high
scanned files  4

[CRITICAL] GHA001
pull_request_target checks out
untrusted pull request code

AI security · Python

Agentic Risk Scan

A dependency-free scanner for AI agent workflows, MCP configurations, and instruction-file risks, with SARIF and GitHub-native reporting.

  • Static analysis
  • GitHub Actions
  • Supply chain

Systems · C · RISC-V

xv6-riscv

A lab workspace for operating-system internals, kernel experiments, local user programs, and the MIT xv6 toolchain.

  • Operating systems
  • Kernel
  • QEMU

Developer tool · Rust

m2html

A fast Markdown-to-HTML renderer with a Rust core, a batch-friendly CLI, Python bindings, and safe defaults.

  • Rust
  • PyO3
  • CLI

Computer vision · PyTorch

Portrait Segmentation

A compact U-Net baseline with training and inference pipelines for binary portrait and human segmentation experiments.

  • U-Net
  • Segmentation
  • Inference

3D vision · Web

3DGS Memory Lab

Experiments around 3D Gaussian Splatting, memory behavior, and browser-facing visualization workflows.

  • 3DGS
  • Vue
  • Visualization
Browse all public repositories

02 / Focus

Four connected tracks.

Learning moves between low-level mechanics and user-facing tools. Each track sharpens a different part of the same engineering practice.

01

Systems

Operating systems, C/C++, RISC-V, digital design, and the mechanics below application code.

02

AI agents

Agent security, memory workflows, instruction transparency, and safer automation boundaries.

03

Vision

Segmentation, pose estimation, 3D Gaussian Splatting, and visual model experimentation.

04

Learning systems

Small tools, practical archives, and reproducible notes that turn coursework into reusable knowledge.

03 / About

Learning in public.

I am a student at Wuhan University, turning coursework and experiments into cleaner repositories with useful documentation, focused scopes, and evidence that the work actually runs.

I am most interested in the boundary between understanding a system and making it useful: tracing an operating system path, testing an AI workflow for unsafe assumptions, or reducing a vision experiment to a reproducible baseline.

The goal is a body of work that reads as both a learning record and a technical portfolio.

Python C C++ Rust JavaScript Vue PyTorch Linux

04 / Contact

Let’s compare notes.

Open to thoughtful conversations around systems, agent security, computer vision, and useful developer tools.

Email Jerry GitHub profile