技術主管,博士研究LLM如何影響人的推理與自主性。 Technical lead, PhD researching how LLMs affect human reasoning and autonomy.
我在做的事,大概可以用一句話概括:當AI產出的速度超過人親自檢查的速度,你要相信誰,這幾年我一直在用不同的專案回答這個問題,現在也把它變成正式的博士研究。 What I do comes down to one question: once AI produces things faster than anyone can personally check them, who do you trust? I've spent the last few years answering that through different projects, and now I've turned it into formal PhD research.
自我陳述不算數,獨立驗證才算。 Self-report doesn't count. Independent verification does.
Legwork、Syco Eval、The Crucible 分別在「程式碼審查」「AI是否諂媚使用者」「單一模型的判斷可不可信」這三個情境裡,實作同一個判準。 Legwork, Syco Eval, and The Crucible apply the same test to three situations: code review, whether AI flatters its users, and whether a single model's judgment can be trusted.
一個純機械式的PR審查輔助工具。不靠AI判斷程式碼好不好,靠覆蓋率疊圖、依賴地圖、影響範圍這些查得到來源的東西,誠實標出「哪裡有人真的驗證過,哪裡沒有」。 A purely mechanical PR review aid. Instead of trusting AI's judgment of whether code is good, it uses coverage overlays, dependency maps, and blast radius — things you can trace back to a source — to honestly mark where something has actually been verified, and where it hasn't.
查看專案View project建立在 Syco Eval 開源研究上的稽核服務,判斷AI輸出有沒有討好使用者甚於講真話。用獨立訓練的分類器去查,不是讓AI自己監控自己的諂媚程度。 An audit service built on the open-source Syco Eval research, checking whether an AI's output flatters the user more than it tells the truth — checked with an independently trained classifier, not by letting the AI monitor its own sycophancy.
查看專案View project開源的多模型辯論工具。與其相信單一模型講得多有把握,不如讓幾個互不知道彼此存在的模型跑同一個判斷,意見分歧的地方,才是真正該讓人看一眼的地方。 An open-source multi-model debate tool. Rather than trust how confident a single model sounds, run the same judgment through several models that don't know of each other's existence — wherever they disagree is exactly where a human should look.
查看專案View project↳ 這裡的分歧偵測技術,已標記為Legwork下一階段驗證層的候選基礎設施。 ↳ The disagreement-detection technique here has been flagged as candidate infrastructure for Legwork's next verification layer.
不屬於主線,各自有自己存在的理由,不需要硬掛上同一個故事。 Not part of the main thread — each has its own reason to exist, with no need to force it into the same story.
幫財富管理從業人員追蹤ELN類結構型商品的工具,小規模使用中。 A tool that helps wealth managers track ELN-style structured products, currently in small-scale use.
中世紀劍與魔法、PvE撤離型roguelike,支援第一/第三人稱。 A medieval sword-and-sorcery, PvE extraction roguelike supporting both first- and third-person.
完整履歷跟聯絡方式在下面 — 往下捲動就會看到,不用照順序讀。 Full résumé and contact info are below — scroll down, no need to read in order.
履歷 — 參考用資料Résumé — for reference
TAO Digital Solutions
帶領台北工程團隊,負責跨境資料交易平台的後端與架構,服務橫跨巴西與美國市場。 Leading the Taipei engineering team, responsible for the backend and architecture of a cross-border data-trading platform serving both Brazil and the US.
國立陽明交通大學National Yang Ming Chiao Tung University
研究LLM如何影響人的推理、批判思考與決策,以及這如何連動到著作權與自主性的喪失,並把研究成果應用在正式的RAG與多模型協作系統上。 Researching how LLMs affect human reasoning, critical thinking, and decision-making, and how that connects to the erosion of authorship and autonomy — applying the findings to production RAG and multi-model collaboration systems.
Innova Solutions
從工程師升任技術主管,主導美國醫療資料平台與跨境交易平台兩個客戶專案的架構與交付。 Promoted from engineer to technical lead, driving architecture and delivery for two client projects: a US healthcare data platform and a cross-border trading platform.
私人教育Private education
全職教授高階數學與物理,2022年轉職進入軟體工程。 Taught advanced math and physics full-time; moved into software engineering in 2022.
AWS Certified Machine Learning – Associate
技能。Skills. Distributed Systems、AWS、RAG / LLM Orchestration。 Distributed systems, AWS, RAG / LLM orchestration.
目前在做工程/研究相關的機會,或單純想聊聊上面任何一個專案,歡迎聯絡。 Currently open to engineering/research opportunities, or just want to talk about any of the projects above — feel free to reach out.