A notebook, in public.
AI · Nanjing University

PERSONAL INDEX GitHub

02 / CODE & EXPERIMENTS

Projects

Things I build, experiments still taking shape, and the notes behind them.

01Selected work

06

Reinforcement learning

stickman-rl-lab

A 2D stickman learns to reach a target with PPO through an obstacle curriculum, with a desktop console for live training.

Full-course deterministic evaluation: 54/80 successes (67.5%). Natural upright walking is not claimed.

Python / PyTorch / Gymnasium / PyMunk

An articulated stickman and live PPO metrics in the desktop training console.
Training console · smoke run

Interactive learning

ai-anomaly-bureau

A browser investigation game about supervised learning, overfitting, hidden tests, and controlled experiments. Models train in the browser; test labels stay hidden until an audit.

React / TypeScript / Vite

A pixel-art cat is misclassified as bread in the opening scene of AI Anomaly Bureau.
Browser game · opening scene

Agent tooling

dsh-mcp-gateway

Connects ChatGPT Web to a local DSH runtime over OAuth and MCP, exposing its existing tools, skills, and policies without wrapping each extension.

Python / MCP / OAuth

Mathematical animation

neural-choreography

A continuous animation of neural-network mechanisms following Dive into Deep Learning: 49 segments, 11 min 21 s at 1080p60.

64 automated checks cover the underlying mathematics and rendering.

Python / Zanim / FFmpeg

Inputs, hidden ReLU neurons, and outputs in an animated multilayer perceptron.
Neural-network animation · still

Geometry tools

deformation-aware-retopology

Quad edge loops for long, tube-like meshes, with bend previews and a Blender add-on.

Branching shapes such as whole hands and bodies are not yet handled.

Python / Blender

Community tools

funuono

A QQ group bot built on AstrBot and NapCat: join-request verification and reaction-vote moderation. Merged pull requests deploy after CI passes, with rollback on failure.

The join-request verifier only auto-approves; it never automatically rejects a request.

Python / AstrBot / SQLite

02In progress

02

Experimental infrastructure

concept2motion

Local patches, fast previews, and evidence-based review for AI-authored mathematical and machine-learning animations.

A stable improvement in expressive quality has not yet been shown.

Planning & bounded validation

labcouncil

A planned platform where AI agents investigate a researcher’s ideas and meet with the researcher to decide the next experiment.

There is no platform application yet.

03Study notes

02

Ongoing notes and exercises on GitHub. Finished essays live in Writing →.

  • learning-notes

    Reading notes and exercises on machine learning and deep learning.

  • nju-class

    Nanjing University course notes: mathematical analysis, advanced algebra, discrete mathematics, and programming.

Education, research & full project details in my CV →