17 年跨產業營運管理、扛過 2.3 億 P&L、領導 200 人團隊。2026 起自學技術,獨立設計並上線一整套真實營運的 AI 產品。 17 years of cross-industry operations leadership — owning a NT$230M P&L and leading 200-person teams. Self-taught engineer since 2026, shipping a full suite of AI products that run in production.
我不是科班工程師。我是帶過 200 人、扛過 2.3 億營收的營運主管——2026 年開始自學寫程式,把現場最痛的問題一個個做成會自己運轉的 AI 系統。我懂的不只是怎麼寫,而是「該做什麼、為誰做、做了會賺錢嗎」。 I'm not a CS grad. I'm an operator who led 200-person teams and owned a NT$230M P&L — then taught myself to code in 2026 and turned my hardest operational problems into AI systems that run themselves. I don't just ship code; I know what to build, for whom, and whether it makes money.
17 年現場帶人、管 P&L、跑 SOP。我做的每個產品都來自親身踩過的痛點,不做空中樓閣的玩具。17 years on the floor — leading people, owning P&L, running SOPs. Every product I build comes from a pain I've personally lived, not a toy demo.
Python / FastAPI 後端、TypeScript / Astro 前端、Docker、Zeabur / Vercel 部署,加 LLM 整合與 OAuth 閉環——獨立把產品送到真實用戶手上。Python/FastAPI, TypeScript/Astro, Docker, Zeabur/Vercel deploys, LLM integration and OAuth loops — I ship products to real users solo.
用 orchestrator-worker 範式指揮多代理、自架 LLM 閘道統一路由、設計人在迴圈的審核閉環。把 AI 當團隊在管,不是當聊天框。I direct multi-agent orchestrator-worker pipelines, run a self-hosted LLM gateway, and design human-in-the-loop approval flows. I manage AI like a team, not a chatbox.
全部由我獨立設計、開發並上線——從產品定位、後端 API、AI 整合到部署維運。以下為其中代表作。All designed, built and shipped solo — from positioning to backend, AI integration and deployment. A selection below.
完整 SaaS 機器:商家綁定 Google 商家檔案後,AI 自動草擬評論回覆,經 LINE 人在迴圈審核閉環後發佈。FastAPI + Google OAuth + Gemini,已上線生產(reviewpulse.lurnatic.com),125 測試綠。重定位為「月子中心口碑→訂房獲客 OS」鎖定灘頭堡市場。A full SaaS engine: merchants link their Google Business Profile, AI drafts review replies, and a LINE human-in-the-loop approval flow publishes them. FastAPI + Google OAuth + Gemini, live in production with 125 passing tests. Repositioned as a "reputation-to-bookings OS" for postpartum care centers.
為 Review Pulse 供血的找客管線:Playwright 持久化爬取 Google Maps 高評論量商家,抓店資料+評論樣本入庫,三層 email 驗證閘門,附自建「找客審閱駕駛艙」。已收 83 家月子中心名單。The lead pipeline feeding Review Pulse: a persistent Playwright crawler harvesting high-review Google Maps merchants, storing profiles and review samples, with a three-tier email-verification gate and a custom review cockpit. 83 leads collected.
對外公文觸達 OS:冷信/回信按緊急度自動分流,手機 Telegram 一鍵核准後經 SMTP 真寄,launchd 排程自動發送+退信熔斷,收信自動判意向擬稿。141 測試綠。An outbound-comms OS: cold emails and replies auto-routed by urgency, approved with one Telegram tap, sent via SMTP with scheduled auto-send, bounce circuit-breakers and AI intent detection on replies. 141 passing tests.
自架 LiteLLM 閘道,11 個模型統一路由(含 Anthropic 原生 Sonnet),作為 AI 團隊「內容量產組」的批次推理引擎,把大批量純文字工作下放免費模型省成本。A self-hosted LiteLLM gateway unifying 11 models behind one router (including native Anthropic Sonnet), powering the batch-inference layer of my AI agent org and offloading bulk text work to free models.
把 Claude Code CLI 接上 Telegram,從手機即時下指令操控開發機;再延伸出統一登入閘道(admin.lurnatic.com),一次授權守護全部內部服務。Wires the Claude Code CLI to Telegram so I can drive my dev box from my phone — extended into a unified auth gate (admin hub) protecting every internal service with one approval.
通用考試刷題 App,PMP 503 題互動訓練、弱點分析、限時模擬考。Expo + Vercel + Turso 全棧,已過 App Store 審核正式上架。A cross-exam quiz app with 503 interactive PMP questions, weakness analytics and timed mocks. Built full-stack on Expo + Vercel + Turso, shipped and live on the App Store.
從基層做到旗艦店經理,再從營運轉向 AI 產品——一條把現場 know-how 變成程式碼的路。From crew to flagship manager, then from operations into AI products — a path turning floor know-how into code.
7 年全英語環境獨立作戰,跨國團隊管理與高階主管協作皆以英文進行。 7 years operating solo in an all-English environment, leading multinational teams and working with senior execs in English.