Qichang Zheng郑其昌 郑其昌Qichang Zheng
I build the environment AI agents run in — and a platform for developers to orchestrate many agents working together on complex projects. 我为 AI Agent 构建运行环境，也为开发者打造在复杂项目中统筹多个 Agent 协同工作的平台。
AI Agent Platform Engineer, based in Shanghai AI Agent 平台核心开发者，现居上海
About关于
At Qiji Zhifeng in Shanghai I lead the architecture and core development of an AI agent cloud platform, and a team of 10. It is where agents are built, deployed, and run, and it is in production across several deployments. 我在上海奇绩智峰主导 AI Agent 云平台的架构设计与核心开发，并带一支 10 人团队。平台支持构建、部署和运行 AI Agent，已在多个生产环境上线。
I also built a distributed multi-agent collaboration platform on my own. All of my agents are scheduled and work together on it, and it is now open to my teammates, each with a hosted virtual machine. 我还独立开发了一个分布式的多 Agent 协作平台，我的所有 Agent 都在这个平台上调度协作；平台已开放给团队成员使用，每人一台托管虚拟机。
With it, in the past month: about 1.83 million new lines of code including tests, 69 features shipped, over 138 billion tokens. 近一月借助它新增代码约 183 万行（含测试），上线新功能 69 项，消耗超 1,380 亿 token。
Before that I was an AI Scientist at Devz AI in the US, working on enterprise agents, a general-purpose web agent, and fine-tuned domain LLMs. 在此之前，我在美国 Devz AI 担任 AI Scientist，做企业级 Agent、通用网页智能体和垂类大模型微调。
Work作品
North Agent Cloud
A platform for building, deploying, and running AI agents.构建、部署和运行 AI Agent 的平台。
- Scheduling, sandbox isolation, egress control, credential custody, elastic scaling, end-to-end tracing; it can be deployed fully offline and on-premises, including on Chinese ARM servers.调度编排、沙箱隔离、出网管控、凭证托管、弹性伸缩、全链路追踪；支持离线私有化部署，含国产 ARM 服务器。
- Runtime and sandbox are decoupled: long, low-load LLM calls and short, high-load tool execution scale separately. Lazy start and automatic reclamation fit about 10x more sandboxes on the same resources.运行时与沙箱分离：长时低负载的模型调用和短时高负载的工具执行各自伸缩；懒启动加自动回收，让同样的资源承载约 10 倍的沙箱。
- The platform’s own built-in features — trajectory analysis, building agents by conversation — are themselves agents running on it.平台内置的轨迹分析、对话式构建 Agent 等功能，本身就是跑在这个平台上的 Agent。
North Sandbox
A multi-tenant, E2B-compatible sandbox service, also offered on its own: point the official SDK at it and the core flows run.兼容 E2B 协议的多租户沙箱服务：官方 SDK 换地址和凭证即可跑通核心流程，也可独立对外提供。
- Per-template warm pools cut acquisition to milliseconds; custom templates build on a base image with your own dependencies.按模板预热，沙箱认领降到毫秒级；支持自定义模板：从基础镜像派生、装上自己的依赖。
- JuiceFS mounts object storage as the working directory: uploads and file views never start the sandbox, and a paused sandbox keeps its files.以 JuiceFS 将对象存储挂载为沙箱工作目录：上传、查看文件直连 JuiceFS，无需启动沙箱；暂停后恢复，文件状态都在。
- Network policy: a new sandbox’s egress can never exceed what its project allows; deny by default.网络策略：建沙箱时校验其出网范围不超过项目允许的上限，默认拒绝、按需放行。
- Usage metering: running time and storage per sandbox, closed out daily for billing and quotas.用量计量：分段记录每个沙箱的运行时长与存储用量，每日封账，为计费与配额提供依据。
North Firewall
An egress firewall built for agents that also runs on its own as an egress proxy gateway for other apps.专为 Agent 设计的出网防火墙，也能单独部署，作为通用的出网代理网关给其他应用使用。
- A transparent, chained forward proxy: sandbox traffic reaches it with no agent configuration; HTTPS is decrypted with a self-signed CA and audited per request, and an upstream proxy can be chained in.透明的链式正向代理：沙箱的出网流量被透明转给防火墙，Agent 不用改任何配置；用自签 CA 对 HTTPS 做 MITM 解析，逐个请求审计，还能再串接上游代理。
- Fine-grained egress policy: allow and deny rules per project and sandbox by domain (with wildcards), IP range and port; deny by default, a multi-tenant console, and a reason code on every refusal.细粒度出网策略：按项目、按沙箱分别配置域名（含通配）、IP 网段和端口的放行与拒绝规则，默认拒绝，带多租户管理台；被拒的请求返回明确的原因码。
- Credential isolation with Vault: agents only ever see placeholders; the gateway swaps in the real key once policy allows the request, so real keys never enter the sandbox.凭证隔离 Vault：凭证以占位符注入，Agent 只接触占位符；策略放行后才在网关里换成真实密钥，真密钥不进沙箱。
Agent Team
A multi-user, distributed platform for agent collaboration: hosted machines, scheduling, and verification in one place.多用户的分布式 Agent 协作平台：托管机器、调度和验证一体。
- Every user gets a virtual machine hosted by the platform. Agents are scheduled there, move across machines, sleep when idle and wake when a message arrives; you can talk to any agent from the web, the mobile app or Feishu, and the platform routes messages between agents. Over the past month, a median of 18 agents ran each day, 20 at peak.每位用户一台平台托管的虚拟机，Agent 在里面调度、跨机器迁移、闲置自动休眠、来消息自动唤醒；网页、手机 App 和飞书都能直接和某个 Agent 对话，Agent 之间由平台路由通讯；近一月每天运行 Agent 中位 18 个、峰值 20 个。
- Minute-level patrol and custom Stop hooks catch the failure modes of long-running work, like an agent that says it will keep going and then stops. Nothing counts as done until a second agent has tried to prove it broken.分钟级巡检和自研 Stop hook 治理长程任务的典型失效——比如说了要继续却停下；每项改动都要经另一个 Agent 端到端对抗验证，证伪不了才算完成。
- 3,900+ commits, about 1.08 million lines. Since it went live, my merged pull requests went from about 41 to about 272 a month, and features shipped from 14 to 67.3900+ 次提交、约 108 万行；上线后我个人月均合入 PR 从约 41 个增至约 272 个，交付功能从 14 项增至 67 项。
- Found that macOS VMs never gave freed memory back to the host, and fixed it in libkrun (2.8k stars); merged upstream.定位 macOS 虚拟机内存不归还宿主的问题，为 libkrun（2.8k star）实现空闲页归还和批量回收，已合入上游。
Experience经历
- 2026.03 — now2026.03 — 至今 Qiji Zhifeng奇绩智峰 AI Tech LeadAI 技术负责人
- 2024.08 — 2026.03 Devz AI (US)Devz AI（美国） AI Scientist
Earlier: Prudential Financial, Machine Learning Engineer (Capstone), and Shannon Investment, NLP Engineer, 2023–2024. 更早：Prudential Financial 机器学习工程师（Capstone 项目）、Shannon Investment NLP 工程师，2023–2024。
Education教育
- 2022 — 2024 University of Chicago芝加哥大学 MA in Computational Social Science计算社会科学硕士
- 2018 — 2022 Xi’an Jiaotong-Liverpool University西交利物浦大学 BSc in Economics and Finance经济与金融学士
Get in touch联系我
Shanghai, China中国上海, ， local time 当地时间