⭐ Code Less, Architect More 🚀
|
Hello. I'm an incoming third-year Software Engineering student at North University of China. 💻(￣▽￣) This summer: Java microservices & middleware, LLM internals, and Spec-driven Coding under Harness Engineering. The honeymoon with vibe coding is over — I care about architecture, maintainability, and code built to last. ⚡(¬‿¬) Fewer throwaway lines, more judgment and workflow. What I'm practicing:
Outside tech I'm a cyclist 🚲 and follow bike racing — the Grand Tours and spring classics. I write a blog, tend a digital garden, track anime seasons, and keep a bit of warmth outside the screen that I don't want to lose. |
|
Coding agents are a fleet, not a single tool. Grok scouts real-time topics on the web (and I enjoy tinkering with Grokbot); GPT goes deep on research and requirement breakdowns; previously used Claude Code routed to V4-Flash alongside MiniMax multimodal generation — core development has shifted to ZCode running on the GLM Lite plan with GLM 5.3 Flash, taking full advantage of its Harness ecosystem and plugin marketplace; OpenCode and Pi for lightweight tasks — lightweight routines are now handed over to Gemini 3.8 Flash inside Antigravity; Cursor handles localized patches and quick inline refinements. I avoid reinventing wheels: configure each CLI with its matching Harness ecosystem and plugin registry, and maintain my custom harness-sync skill to synchronize and iterate cross-tool configurations in one click. The harness matters more than the prompt: scoped context, repository rules, deterministic commands, tests, clean docs, and a careful final diff review. 📋(・ω・)ノ Architectural judgment and every merge stay with me.
|
Frontend & Fullstack Java / Spring Middleware & microservices |
AI / agents Systems / inference — competition proof Python DevOps |
|
|
2026 全国大学生计算机系统能力大赛 · 智能计算创新设计赛（先导杯）赛题 1 · Team 翻斗花园 Proof split: official best anchor is 87.7839 / 100 · #26 / 132 · SLA 0 · precision 0; our own best anchor is 87.6933 / 100, with 4k–8k / 8k–16k / 16k–32k throughput 20.39 / 18.29 / 14.61 tok/s and precision 0. Workload: raise long-context Qwen throughput on a fixed domestic DCU at concurrency=1, under TTFT/TPOT P99 SLA (input mix 4k–8k / 8k–16k / 16k–32k). Stack: vLLM 0.18.1 · Qwen3.5-27B BF16 · Hygon DCU (gfx936) on SCNet. My focus: shared-gate fusion, SwiGLU HIP kernels, GDN launch packing, Gather-FA routing, LPK prefetch. Validated recipe: fused shared-gate ON · LPK stages=1 · Gather-FA ≤16K · rocBLAS + TunableOp · LPK prefetch. LONG prefill and LLMM1+fusion were rejected after measured regressions, including 83.8886 for the latter. vs official baseline smoke: TTFT P99 −61%–87% · TPOT P99 ≈ −35% · throughput +7%–24%. Scores locked to SCNet runs, not local Windows numbers. |
2026 书生国智科探挑战赛暨飞翔杯 AI Agent/Skills 开发大赛 · Shanghai AI Lab × Biren · Track 5 模型与算子 · Team 翻斗花园 Live board: public NS64 rel-L2 0.035115 · tag Problem: ship a Biren-native Spectral Convolution (SUPA / Extension) and reuse it inside ≥4-layer FNO-NS on the public 64×64 Navier–Stokes set (1000/128) — Agent/Skills log required (~15%). Stack: Biren106B · SDK 1.11 · Evidence: formal Spectral idle frozen; chain CPU↔SUPA <1e-4; Pred/GT viz on public NS64; Agent log 35+ reviewed segments with abort/NO_SIGNAL discipline (no silent promote). vs prior formal v8 (0.035302): public L2 ≈ +0.53% error drop after dual-view consistency polish; Spectral ms unchanged on purpose. |
|
|
|
|
|
|
|
|
|
Helps DAO operators prepare and approve treasury payouts through Cobo Agentic Wallet, instead of ad-hoc spreadsheets and opaque transfers. Hackathon prototype.
|
|
DSH plugin ecosystem: Diving into the DSH plugin architecture — contributing one plugin to the official community while independently developing another from scratch.
Distributed delay-delivery service: Spring Cloud & middleware practice with high reliability and load-leveling focus.
Enterprise Dify customization: Tailored knowledge-base retrieval and agent workflow extensions for production scenarios.
Grokbot experiments: Tinkering with real-time agent interactions, scheduled monitors, and automated workflow playbooks.
- Java microservices full-stack: building clean microservices end to end — Spring Cloud Alibaba, Nacos, distributed transactions, and production-grade backend engineering.
- AI-assisted E2E testing: using MCP tools so models can drive end-to-end flows and catch broken business paths early. 🔬(・∀・)
- Light CLI + heavy IDE: ZCode / Claude Code for rapid development and exploration, Antigravity (Gemini) for lighter tasks; Cursor (browser, debugger, project context) for long-lived maintenance.
- Paradigm shift: Prompt → Context → Harness → Loop Engineering. Human rules and discipline shrink the human-agent gap and force clearer architecture thinking. 🧭(｀・ω・´)
- Cross-harness sync & custom skills: maintaining multi-environment configs via my custom
harness-syncskill, and distilling practical workflows into the TTA skill fleet. - Indie shipping path: App / mini-program + overseas payments end to end; letting information drive development while filling gaps with AI.
- Spec-driven coding (SDD): practicing Spec-Kit / OpenSpec patterns to align human-agent assumptions through clear upfront contracts.
- Agent engineering (Python): LLM routing, Function Calling / MCP runtimes, stateful Agent Loops with checkpoints, hybrid-retrieval RAG.
- Multi-Agent collaboration & evals: Subagent orchestration, golden datasets, LLM-as-Judge automated evals, tracing, and cost governance.
- Agentic product shipping: moving from idea to prototype (PM → Builder), weaving memory systems, proactive touchpoints, and dynamic UI into real applications.
Last updated: September 2026