✳ Things I've built · ⚡ On my desk · 🎮 Side quests
I'm an AI undergraduate, graduating in 2027. I like turning “what happens if…” into something I can run, break, and improve.
My current rabbit holes: AI agents, reinforcement learning, and robots. Still learning. Still changing my mind. Building along the way.
Six projects, a few rabbit holes. Pick a cover and take a look inside.
📎 Prefer a text-only project index?
| Project | What I'm exploring |
|---|---|
| CarryTrace | Keeping useful context across long-running agent work. |
| Deep Native | Lighter coding-agent behavior, tested instead of assumed. |
| FlowCredit Research | Research memory that keeps the evidence attached. |
| RL Sentinel | Replaying what was known when a recommendation was made. |
| Go2W MoRA | Hierarchical robot navigation in simulation. |
| FlowCredit | Inspectable evidence and deterministic risk signals through an API. |
🤖 Robots that can handle a plan going wrong.
G1 mission experiments: structured tasks, measured outcomes, and failure handling in MuJoCo. Simulation research, not a claim of real-world robot safety.
🧠 An agent that doesn't lose the thread.
CarryTrace: bringing the right decisions and context back when work spans conversations.
🛠️ Doing less, checking better.
Deep Native: experimenting with how much process a coding task actually needs. The first pilot did not show a quality improvement; that result stays visible.
A hand-curated workbench, not a live activity feed. The linked repositories are the source of truth.
Open the other tabs in my brain ↗
Creative coding · AI visuals · Small tools · Interface experiments
Not every idea needs to become a startup. Some are worth building just to see what happens.
我还是学生。想把好奇心变成能跑起来的东西，也想把它们做得好看一点。
🧪 Playful outside. A little more serious under the hood.
I want the experiments to be fun and the claims to be honest.
Keep the failed runs. Link the evidence. Say what isn't working yet.
The project repositories hold the protocols, limitations and results. This profile is a way in, not a substitute for them.
🤖 About the little robot & this page
The robot is an original geometric SVG illustration, not a photo or a product screenshot. It floats, blinks and settles in under five seconds. Motion is decorative; it does not indicate that an agent is online.
This page uses repository-hosted artwork, ordinary links and native disclosure panels. No tracking widgets, external font requests, made-up metrics, or JavaScript pretending to work in a README.
Motion-sensitive readers get a static cover when their browser exposes the reduced-motion preference. Static covers are also available directly: light · dark.
Curiosity first. Keep building. ✳
Eason / Anhao1314 · Learning in public.