_{⭐ Code Less, Architect More 🚀
🤖 Models secure the baseline; experience expands the frontier.}

Hello! I'm threetwoa, a third-year Software Engineering student at North University of China. I mainly use Java, Python, and TypeScript to build desktop tools and personal websites, and turn lessons from AI-assisted development into reusable Skills. I start with my own needs, customize open-source projects, refine interfaces and workflows, and keep maintaining them. Projects & contributions pane · AI usage workbench — A local-first desktop tool I develop and maintain. It lives in the system tray, bringing together usage, quotas, balances, and reset times across AI services and accounts, alongside local CLI spend. Firefly · Personal blog & digital garden — Customized from Firefly / Astro, with reading improvements, site AI Q&A, and portfolio interactions. An agent workflow connects content organization, validation, and publishing. Source TTA Skills — Turning my development practices into agent instructions and checks: tta-init organizes project entry points and context; tta-tone improves writing while preserving source facts; tta-frontend guides frontend tasks, data states, and component boundaries. Outside tech, I ride road bikes, follow Grand Tours and Monuments, and tend my blog and digital garden, documenting the tools I explore and projects I build. 🚲

🌏 Earth Online Player · Builder · 🚲 📍 Taiyuan · NUC  ·  💼 Seeking 2027 SWE Internships 🌱 Core Java / Python Agent Fullstack

                   

---

Projects & highlights

• AgentCFO — A DAO treasury assistant and hackathon finalist prototype. I led frontend development of the landing page and operator console; the team verified two end-to-end settlements on Sepolia testnet. Demo
• Lead Cup · vLLM on Hygon DCU — Worked on long-context inference optimization on a fixed DCU. Our team's own best score was 87.6933 / 100; throughput improved 7%–24% against the official baseline smoke test.
• AI4S · 书生国智科探挑战赛 — Built a Biren-native Spectral Convolution and used it in FNO-NS. The v9 report records public NS64 rel-L2 of 0.035115; final ranking pending.