Currently focused on agent orchestration & LLM tooling.
📍 Shanghai, China
🧭 RAG · Agent Infra · LLM Systems · Decision Models
|
Zhipu AI (智谱) — Nine years across China's largest social platform and its frontier LLM lab — NLP & infrastructure at WeChat, then agent infrastructure at Zhipu (GLM). |
|
- dsh-super-agent — admission control, observe mode & tickets for any CLI coding agent: admission tables, explicit/inferred strength, off → observe → auto, per-ticket worktrees, PR delivery with approval-gated auto-merge, GitHub App identity. Works with
claude/codex/opencode/ anything. - awesome-jev-prompts — question patterns, anti-patterns & calibration notes for typed decisions: 63 patterns / 10 categories, bilingual, an Engine Cookbook that runs one question across five engines, and a live leaderboard of the ecosystem.
- RoboHarness — an incident-hardened harness for $200-class 5-DOF robot arms: a cloud LLM plans (1–3 calls per task), deterministic primitives act, and a real-hardware safety layer — current-spike fuse, z-clearance tiers, contact locks — owns every motion. Zero training; every gate traces to a documented real-robot incident. First fully-automatic closed-loop grasp on real hardware (scan → center → grasp → verify, 2026-11-14). Same thesis as dsh-super-agent, one layer down the stack: the model proposes, deterministic code owns the dangerous parts.
Focus areas: agent infrastructure & admission control, RAG in production, knowledge platforms, decision-layer models, question design & calibration, embodied agents on cheap hardware.
中文:9 年工程经验——腾讯微信 5 年(NLP 与基础设施),智谱 4 年(LLM 工具链与 Agent 基础设施)。现在主推三个项目:dsh-super-agent(DeepSeek Super-agent 准入控制的独立实现,任意 CLI 编码智能体可用,npm 已发布)、awesome-jev-prompts(63 条双语问题模式 + 引擎手册 + 生态排行榜)与 RoboHarness($200 级 5 自由度机械臂的事故加固 harness——云 LLM 规划 + 确定性基元执行 + 电流反馈安全层,零训练,已实现真机全自动闭环抓取);另为 WeKnora 贡献者(15 个被合并 PR;2026 年 9 月单月提交与合并双榜第 4)。