连续创业者。哥伦比亚大学计算机工程，曾任科技创业公司 CTO，打造过北美社交榜 Top 4 的 AI 社交应用。现在我带着 miniCTO——一支 AI 原生交付团队，住进你的业务里，把 AI 变成实打实的产出。 Serial founder. Columbia Computer Engineering, former startup CTO, built an AI social app that hit Top 4 on the U.S. social charts. Now I run miniCTO — an AI-native delivery team that moves into your business and turns AI into real output.
多年产品 + 工程双栖，三段方向不同的创业，练的都是同一件事：把新技术变成别人用得上的东西。 Years across product and engineering, three ventures in three directions — all practicing the same craft: turning new technology into something people actually use.
创办另类数据公司百观（BigOne），把电商、物流、社媒等海量另类数据加工成机构投资人可以直接下单的洞察。在「数据 → 判断」这条链路上，练出了对数据资产和商业语义的手感。Built BigOne, an alternative-data firm — think YipitData for China — refining raw e-commerce, logistics and social signals into insights institutional investors could trade on. That's where the instinct for turning data into judgment was forged.
从零打造 AI 社交应用，冲上北美 App Store 社交榜第 4。这段经历证明了一件事：AI 产品的胜负手不在模型，在体验的分寸感。Built an AI social app from scratch that climbed to Top 4 on the North American App Store social charts. The lesson: AI products are won not on models, but on the finesse of the experience.
All-in AI 原生开发（AIFD），创办 miniCTO：把「建议、开发、运维」折叠成一种可订阅的职能，让每家企业都请得起一支世界级的软件团队。Now all-in on AI-first development (AIFD) with miniCTO: folding advice, development and operations into one subscribable function — so any company can afford a world-class software team.
当所有人都能生成代码和图片，稀缺的是知道什么值得做、做到什么程度算好。miniCTO 交付的每一件东西，都要经得起这句话的审视。When anyone can generate code and images, what stays scarce is knowing what's worth building — and what "good" looks like. Everything miniCTO ships is held to that bar.
所有能力最终都指向这两个结果——先让现在的团队更省、更快；再让公司长出自己的 AI 能力，不再依赖任何外部团队，包括我们。 Everything we do serves these two outcomes — first make your current team leaner and faster; then grow AI capability inside your company, until you no longer depend on any outside team. Including us.
数据采集、报表自动化、流程机器人——把重复劳动交给机器，人去做判断。Data collection, report automation, workflow bots — repetitive work goes to machines; people keep the judgment calls.
海报、素材、落地页的 AI 产线：设计需求从排期几天变成对话几轮。An AI pipeline for posters, assets and landing pages: design requests go from days in a queue to a few turns of chat.
数仓治理、指标口径、BI 看板与专项分析——取数不再等人，决策不再靠猜。Warehouse governance, metric definitions, BI and deep-dives — no more waiting for numbers, no more guessing.
企业内部 AI 助理、对话式产品、多智能体协作——从原型到生产环境。In-house AI copilots, conversational products, multi-agent systems — prototype to production.
小程序、Web、App，从设计稿到上线运营，一支团队闭环走完。Mini-programs, web and mobile apps — design through launch and operations, one team end to end.
像部门一样入驻，边交付边把方法留在你的团队里——AI 化是能力移交，不是外包依赖。We embed like a department, shipping while transferring the playbook to your team — AI transformation as capability handover, not outsourcing dependency.
过去，企业请咨询公司出建议、请外包公司写代码、再雇人维护。AI 把这三件事折叠成了一件：一支足够小、足够快的驻场团队，可以从理解业务开始，一路做到软件上线并持续进化。 Companies used to hire consultants for advice, outsourcers for code, and staff for maintenance. AI collapses all three into one thing: a small, fast resident team that starts from understanding your business and stays through launch and beyond.
我们按职能入驻：懂你的数据、你的流程、你的客户，交付之后不消失。We embed as a function — fluent in your data, your process, your customers. We don't disappear after delivery.
用 AI 把交付周期从季度压到星期，但代码评审、测试与上线流程一个不少。AI compresses quarters into weeks — without skipping review, testing, or release discipline.
AI 干活，真人负责。每个项目都有能拍板、能背责任的人对接你。AI does the work; humans own the outcome. Every project has a person who can decide — and be accountable.
「人不减少，产出十倍。Headcount stays. Output multiplies.」
以下均为真实交付项目，应客户要求隐去名称。从数据平台到移动应用，从政务小程序到 GPU 视频管线——同一支团队，跨行业闭环。All real engagements; client names withheld by request. From data platforms to mobile apps, government mini-programs to GPU video pipelines — one team, closing the loop across industries.
为一家连锁医美 SaaS 平台入驻式服务：治理近 800 张表的数仓、统一指标口径、产出经营专项分析；同时承接小程序营销活动的完整开发上线——从需求拆解到发版，走通了「数据 + 产品」双线交付。Resident engagement with a medical-aesthetics SaaS platform: governed a ~800-table warehouse, unified metric definitions, delivered executive analyses — while also building and shipping mini-program marketing campaigns end to end. Data and product, both lines running.
从公开数据自动采集，到宏观洞察、竞品足迹矩阵与战略沙盘，再到可对话的情报 Agent——一个带登录门户的完整情报系统，让战略团队随时「问数据」。From automated public-data collection to macro insights, competitive-footprint matrices and a strategy sandbox — capped with a conversational intelligence agent. A full gated portal where the strategy team can simply ask the data.
从小程序前端、后端服务到运营管理台的完整承接：预约派单、微信支付与订阅消息、活动与优惠体系。为了不切断老版本用户，新旧两套后端在生产同机并行分流运行；服务过程的视频日记接入 AI 解析，自动归位到对应宠物与日期，人工只做终审。End to end: mini-program client, backend services, ops console — booking and dispatch, WeChat Pay and subscription messages, campaigns and coupons. Two backend generations run side by side in production behind traffic splitting, so legacy app versions never break. Service video diaries flow through an AI parser that files each clip to the right pet and date; humans only sign off.
Flutter 移动应用（跟读评分、词汇闯关）、内容与用户管理后台、招生小程序、品牌官网、AI 生图工作站与网页课件，构成一个完整的品牌矩阵。同时打通微信与支付宝双通道支付，并走完从域名备案到应用商店上架的全部合规链路。A Flutter app (shadowing with pronunciation scoring, vocabulary quests), a content and user admin, an enrolment mini-program, the brand site, an AI image studio and browser-based courseware — one full brand stack. Dual-rail payments across WeChat and Alipay, plus the entire compliance path from domain filing to app-store release.
接入微信自动续费签约与免密扣款、打通广告平台的融合归因回传，让投放花的每一分钱都能追到订阅收入；同时为海外市场拉起独立部署分支。研究侧在 8 卡 GPU 机上跑音频理解与自动 MV 生成的长链路流水线。Wired up WeChat recurring-payment contracts and silent billing, plus blended attribution callbacks to the ad platform so every yuan of spend traces to subscription revenue. Stood up a separate overseas deployment branch, while the research side runs a long-chain audio-understanding and auto-MV pipeline on an 8-GPU box.
编排十余个文生图 / 图生视频 / 语音模型，把「设定 → 剧本 → 分镜 → 成片」做成一条可重试、可断点续跑的串行链路；尾帧串联保证镜头之间人物一致，互动剧分支编辑器自带可达性与断链校验。Web 端、小程序、内容后台与短剧 App 之间打通了完整的发布链路。Orchestrates a dozen text-to-image, image-to-video and voice models into a resumable serial chain: premise → script → storyboard → finished cut. Last-frame chaining keeps characters consistent across shots; the branching editor validates reachability and dead ends. Web, mini-program, content console and the drama app share one publishing pipeline.
同一客户的两条线：口腔器械电商（商品、订单、后台）与面向医生的临床 RAG 问答助手——中文向量检索 + 推理模型流式作答，把分散的临床资料变成可追溯出处的回答。含邀请码门禁与生产环境部署运维。Two lines for one client: a dental-device storefront (catalog, orders, admin) and a clinical RAG assistant for practitioners — Chinese vector retrieval plus a reasoning model streaming answers, turning scattered clinical material into citable responses. Invite-gated, deployed and operated in production.
官方文旅小程序 + 内容管理后台：景区、线路、活动的内容中台，首页多套皮肤后台可切；内置 AI 助手做对话问答与行程规划。交易类能力按分期规划预留数据结构，一期先把内容与 AI 体验做扎实。An official tourism mini-program plus a content console: a CMS for sights, routes and events, with switchable home-page skins; an embedded AI assistant handles Q&A and itinerary planning. Commerce is scoped to a later phase, with the data model reserved up front.
接手第三方交付的联盟链数藏平台源码：审计出硬编码密钥与生产数据泄露风险并给出脱敏方案；把原本依赖图形化 IDE 的发版流程改造成一条命令行 CI 通道，从此发版不再依赖某台特定的电脑。Inherited a third-party consortium-chain collectibles codebase: audited it, surfaced hard-coded keys and production-data exposure, and laid out remediation. Replaced a GUI-IDE-dependent release ritual with a one-command CI pipeline — shipping no longer depends on one particular laptop.
为酒店集团搭建 OTA 平台经营指标的每小时自动采集与入库，异常自愈、失效告警，运营团队从手工抄数中彻底解放。Hourly automated collection of OTA operating metrics for a hotel group — self-healing, alert-on-failure. The ops team never copies numbers by hand again.
为客户部署专属 AI 设计师 Agent：网页端对话协作、海报与营销素材生成、素材资产管理。设计需求从排期数天变成对话几轮。Deployed a dedicated AI-designer agent: web-based chat collaboration, poster and marketing-asset generation, asset management. Design requests went from days in a queue to a few turns of conversation.
AI 生成测试计划、真人测试员实机执行、结构化报告自动回传——把「人工验收」做成了一个可调用的服务。AI writes the test plan, human testers execute on real devices, structured reports flow back automatically — manual QA, turned into a callable service.
我们持续用自研产品验证新技术，跑通了再带进客户项目——客户不为我们的学习曲线买单。We prove new tech on our own products first, then bring the working parts into client work. Clients don't pay for our learning curve.
分割 + 修复模型组合，做到画面人物擦除、指定人物保留、其余风格化，以及稀疏关键帧驱动的形象编辑器——比逐帧估计更稳，也更省算力。Segmentation plus inpainting models: erase people from footage, keep a chosen subject, stylize the rest — driven by a sparse-keyframe editor that beats per-frame estimation on both stability and cost.
常驻飞书与 Slack 的一组内部 Agent：口头一句话就能派单建任务并追踪状态、表格变更自动播报、多租户机器人按频道隔离工作区。管理动作发生在对话里，不在又一个后台。A set of agents resident in Feishu and Slack: dispatch a task by saying it out loud and have it tracked, broadcast spreadsheet changes automatically, isolate workspaces per channel in a multi-tenant bot. Management happens in conversation, not in yet another dashboard.
语言朗读训练、创作者增长站、创意投票社区、行业情报雷达、互动小游戏合集……用最小成本验证新玩法，再把验证过的能力沉淀成可复用的模块。A speaking trainer, a creator-growth site, an idea-voting community, an intelligence radar, a hub of playable demos… smallest-cost validation of new mechanics, distilled into reusable modules.
下面每一项都在真实交付里跑过生产，不是简历上的关键词列表。遇到没用过的，我们学——但不拿客户的项目当练习场。Everything below has carried production traffic in a real engagement — not a keyword list on a résumé. When something is new to us, we learn it — just not on your project.
一次深谈 + 快速调研，找到 AI 能撬动的最大杠杆点，给出明确的第一步。One deep conversation plus rapid research — find the biggest lever AI can pull, and name the first move.
小团队驻场（远程或现场），几周内交付第一个可用版本，用结果建立信任。A small team embeds — remote or on-site — and ships the first working version in weeks. Trust is built on results.
按职能持续订阅，系统随业务一起进化。走得深的伙伴，我们也谈技术入股。Subscribe to us as an ongoing function; the system evolves with the business. With the right partners, we also take equity for technology.
带着它来。一杯咖啡的时间，我们告诉你 AI 时代它应该怎么做、多快能做出来。Bring it over. In the time it takes to finish a coffee, we'll tell you how it should be built in the AI era — and how fast.