我把 AI 做成真正在用的工具——不是 demo，是上線跑的系統。金融分析、地方創生、學術查核，每一個都從真實痛點出發。 I build AI tools that actually run in production — not demos. From quantitative finance to community revitalization, every project starts from a real problem.
三版迭代Three Iterations
- V1 永續循環：執行永續行為 → 點數 → 企業ESG數據回流V1 Circular: Actions → points → corporate ESG data loop
- V2 文化科技：LLM+ASR 將耆老口述轉為像素RPG資產V2 Cultural Tech: LLM+ASR converts elder oral history into pixel RPG assets
- V3 校園版：LINE-based行為驗證＋實體Dashboard，已與信義書院合作V3 Campus: LINE verification + physical dashboard, live with Shinsyi Academy
薪火計畫 — 阿嬤記憶公仔GaiaCraft Spark — Memory Doll
深坑炮仔崙的90歲阿嬤是整座山的活記憶庫。我們用AI記錄她的聲音，3D列印她的樣貌，製成會說話的記憶公仔——擔任景點導覽與無人商店引導，讓記憶在薪盡前留存。A 90-year-old grandma in Shenkeng holds the living memory of an entire hill. We captured her voice with AI and 3D-printed her likeness into a talking memory doll — for tourist guidance and unmanned store interaction, before the memory fades.
USR數位孿生地圖USR Digital Twin Maps
永安、深坑兩個USR場域製作像素地圖，遊客可在虛擬空間走動探索在地文史，實現數位孿生的地方創生。Pixel maps of Yong'an and Shenkeng USR sites let visitors explore local history in virtual space — digital twin community revitalization.
影片展示Demo Videos
成果Results
教育部人文永續 決賽MOE Humanities Finals 中鼎基金會 決賽CTCI Foundation Finals 政大AI競賽 佳作NCCU AI Merit Award 國科會計畫 獲選NSTC Grant Selected核心架構亮點Architecture highlights
- 宏觀篩檢：VIX、DXY、Fed利率 → Regime Score (0–100)，動態調整曝險上限Macro filter: VIX, DXY, Fed rate → Regime Score (0–100), dynamically caps exposure
- AI基本面：Claude API + 12宮格決策矩陣 + 強制Bear Case論述AI fundamental: Claude API + 12-cell decision matrix + mandatory bear case
- Kill Questions：三個殺手問題強制對抗確認偏誤Kill Questions: 3 forced questions combat confirmation bias every trade
- ATR部位管理：依價值/成長/投機分級風控，相關性矩陣防曝險集中ATR position sizing: tiered by value/growth/speculative, correlation matrix prevents concentration
- 「尊重圖形」原則：技術面與基本面衝突時技術面優先，Telegram推送決策卡片Chart-first rule: technical signals override fundamentals; Telegram decision cards preserve human judgment
研究報告知識庫（Layer 1B）Research Report Knowledge Base (Layer 1B)
整合EMBA學長提供的付費券商研報，AI自動分類摘要建立索引，注入Claude分析prompt以提升消息面準確度與深度。Integrates paid institutional reports from EMBA alumni. AI classifies and indexes each report, injecting summaries into Claude's analysis for institutional-grade edge.
回測驗證 → 60天畢業條件Backtest → 60-day paper trading graduation
vectorbt引擎，70/30切分，Sharpe>1.0、Max DD<20%方可上線。Paper Trading 60天連續穩定後轉真實交易。vectorbt engine, 70/30 split; Sharpe >1.0 and Max DD <20% required before live. 60-day paper trading graduation condition.
指導：政大AI中心陳昭伶老師。生醫文獻中39.18%的引用存在錯誤。系統透過GROBID萃取、動態微型向量庫，Citing RAG（chunk=1500）＋ Reference RAG（chunk=1000），LLM裁判比對後輸出Excel查核報告。Gemma & Qwen 72%最高，Claude 60%，Llama 56%。Supervised by Prof. Chen, NCCU AI Center. 39.18% of biomedical citations contain errors. GROBID extraction, dual RAG layers, LLM judge outputs Excel reports. Gemma & Qwen 72%, Claude 60%, Llama 56%.
服務「不急的日常配送」——物流成本最高、效率最低的市場。SAM約28萬人，Phase 1 C2C捷運沿線社區驗證，Phase 2 B2B成為電商最後一哩路夥伴。ESG：綠色里程轉為企業資產，打造共融就業社會。合作：勞動部發展署、銀髮人才資源中心。Targets non-urgent daily delivery — highest cost, lowest efficiency segment. SAM 280k. Phase 1: metro-adjacent C2C; Phase 2: B2B e-commerce last-mile. ESG: green mileage as corporate asset. Partners: MOL, Silver Talent Center.
發現政大公車App即時資訊不準，真正的訣竅是「發車時間+5分鐘才到山上」。獨自開發網頁顯示距當前最近班次，支援加入Safari主畫面成PWA App。一篇Threads貼文引爆校園討論。NCCU's bus app was inaccurate. The real trick: add 5 min to departure time for uphill arrival. Built a page showing nearest departure, with PWA install support. One Threads post went viral campus-wide.
⌥ GitHub → 網站 →Live →以 Google Apps Script 串接 Google Sheets 作為後端資料庫，打造輕量化排班管理解決方案，大幅降低人工行政負擔，實現即時同步的數位化管理。Google Apps Script + Sheets backend builds a lightweight shift management solution, eliminating manual admin work with real-time sync.
為政大游泳隊量身打造的數位化管理平台，以 Google Apps Script 搭配 Google Sheets 作為後端，提供即時同步的隊務管理方案，涵蓋隊員簽到、課表發布、請假審核、訓練日誌四大核心功能。Custom digital management platform for NCCU's swim team. Google Apps Script + Sheets backend provides real-time sync across four core modules: member check-in, schedule publishing, leave approval, and training logs.
系統入口 →Live System →企管、統計、資管，三個系，每一個我都是自己選的。現在回頭看，那些拼起來剛剛好：商業的直覺、數字的嚴謹、程式的執行力，加上金融的視角，讓我可以在不同語言之間自由切換，真正把東西做出來。Business, Statistics, Information Management — three majors, each one I chose myself. Looking back, the pieces fit: business intuition, statistical rigor, engineering execution, and financial perspective. Together, they let me switch between different languages of thinking and actually build things.
我創辦了政大金融科技創新實驗室，邀請政大金融科技研究中心主任王儷玲擔任指導老師；與EMBA學長共同開發Alpha Pipeline十層量化投資系統；帶著沈錡大使獎學金赴日本九州實地考察台積電供應鏈；在2026全國大專國際經貿競賽以供應鏈重組研究拿下全國第二。I founded NCCU's FinTech Innovation Lab, secured Prof. Wang Li-Ling as advisor. Co-developed Alpha Pipeline with EMBA alumni. Flew to Kyushu on scholarship to investigate TSMC's supply chain. Won National 2nd at the 2026 Int'l Trade Competition with supply chain research.
但讓我最難忘的，是在深坑山上遇到那位90歲的阿嬤。她是整座山的活記憶庫。我們把她的聲音用AI留下來，把她的樣貌用3D印出來，讓記憶公仔繼續說話。But what stays with me most: meeting a 90-year-old grandma on a mountain in Shenkeng. She was the living memory of the whole hill. We captured her voice with AI, 3D-printed her likeness, so the memory doll keeps talking — before the fire goes out.
我也游泳，代表政大參加全國大專運動會，自由式接力全國第七。泳道和程式碼教我同樣的事：只要持續推進，就能突破。I also swim — representing NCCU at the National University Games, 7th in freestyle relay. The lane and the codebase teach the same thing: keep moving forward, and you'll break through.