Skip to content
View JamesLeeCY's full-sized avatar

Block or report JamesLeeCY

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
JamesLeeCY/README.md

Chun-Yi Lee | 李俊毅

Data scientist working on high-dimensional, noisy, multimodal data — industrial sensors, finance, language, and signal. 專注於高維度、高雜訊、多模態資料的資料科學家 — 工業感測、金融、語言、訊號。

我建構可以上線、也可以被量測的資料與 AI 系統:引用溯源的 RAG、多代理人研究系統、 中文對話分析,以及正式上線的報名與金流系統。十年處理神經影像與行為資料的訓練, 讓我習慣在高雜訊資料上做嚴謹的評估,現在用在製造、金融與語言資料上。

I build data and AI systems that ship and can be measured: citation-grounded RAG, multi-agent research systems, Chinese conversation analytics, and a live registration and payment platform. Ten years of neuroimaging and behavioral data made rigorous evaluation on noisy data a habit — now applied to manufacturing, financial, and language data.

📍 Taipei, Taiwan | 正在尋找資料科學、機器學習、AI 工程職缺 📍 Open to data science, ML, and AI engineering roles in Taiwan


代表作品 | Selected work

paper-rag-copilot — 引用溯源的 RAG 系統 A strictly citation-grounded RAG assistant with a quantified evaluation harness — and a second harness that checks whether the LLM judges doing the measuring can be trusted.

檢索是簡單的那一半。強制引用的 prompt + 逐字引文檢查 + 經驗證的 LLM 評審(phi4,誤放率 5% [2–15]);章節切塊 Hit@5 100% vs 固定長度 88%;逐則稽核後補上章節標題、允許跨段引文,嚴格引用精確率 84% → 97%。所有比率都附信賴區間,可完全本地運行。 RAG hybrid search prompt & context engineering LLM-as-judge hallucination evaluation Ollama

fmri-video-feature-analysis — 影像特徵萃取與時序對齊 Per-TR visual, motion, semantic, and object-level features from naturalistic video, extracted with OpenCV, HOG, and YOLOv8.

在 0.752 秒的解析度下把原始影像轉成能通過統計檢驗的特徵序列。 computer vision YOLOv8 feature engineering multimodal time-series

line-chat-triage — 中文對話風險分流 Health triage over LINE work groups — six business-time-aware indicators, a non-compensatory escalation tripwire, and automated PDF reports; LINE webhook and Telegram ingestion.

把非結構化的中文對話紀錄,變成可排序、可行動的風險訊號。以程式埋入標準答案的 100 組合成評估:改進後的規則系統在驗證組風險等級正確率 94%、誘餌零誤報;股票社群模式以 entropy 時間序列預測話題轉移(AUC 0.62 → 0.71),LLM 多空標註與人工的 Cohen's κ 最高 0.42 / 0.55。 Chinese NLP semantic entropy LLM labeling synthetic benchmark FastAPI

PACLIC 40 registration — 國際會議線上報名與金流系統(正式站,repo 私有) Registration and payment system for PACLIC 40 (Dec 2026, NTU), live since 2026-10-01: Google sign-in, student-ID review, NewebPay card payments, admin CSV export.

與團隊協作,我負責後端功能與上線維運:查表式計價、截止邏輯、學生證到期自動刪除、Docker Compose + Nginx 部署、CI、排程備份與還原演練。 TypeScript NestJS Next.js PostgreSQL Docker GitHub Actions

semiconductor-quality-ml — 半導體品質與良率預測 Anomaly detection and yield prediction across three public manufacturing datasets — a static snapshot (SECOM), in-process time series (PHM 2016 CMP), and a production routing graph (Bosch).

主張:方法選擇應該跟隨資料結構,而不是跟隨潮流。PHM CMP 官方測試集 MSE 6.24;以診斷把關模型選擇,例如 CMP 虛擬量測 R² 0.993 在主要群集內為 0.897、驗證集最佳的 1D-CNN 在測試集最差——避免把假改善帶進部署。 imbalanced classification concept drift virtual metrology GNN XGBoost — 完整分析 →

precision-weighted-investment-agent — 多代理人精度加權投資研究系統 Multi-agent system (fundamentals, news, macro/flows, supply-chain knowledge graph, technicals) that ports precision weighting from the predictive-processing framework to Taiwan-equity judgments, tested on a survivorship-bias-free universe (271 stocks, 2019–2026, 8,475 walk-forward predictions).

關鍵發現:精度加權的前提是能可靠估計各來源的準確度,實測估計誤差與訊號同量級,因此系統改以財報 Agent 為主。月營收驚奇(SUE)是第一個在 2019–2022 與 2023–2026 兩段都顯著的訊號(產業內 IC +0.032 / +0.059)。財報 Agent 的產業內選股有效(IC +0.029,t 2.16,各子期間與取樣起點皆為正);選股策略的 Sharpe 尚未超越 0050,樣本外效果較小。規則已凍結並事先登錄評估方式,每月排程自動做前瞻驗證。 multi-agent LLM walk-forward backtesting information coefficient survivorship bias deflated Sharpe forward validation — 研究儀表板 →

All projects were built with Claude Code as a development assistant. | 所有專案皆以 Claude Code 協助開發。


技術 | Tools

Python R MATLAB SQL TypeScript JavaScript scikit-learn XGBoost graph neural networks time-series analysis RAG LLM evaluation multi-agent systems Ollama ChromaDB OpenCV YOLOv8 signal processing FastAPI NestJS Next.js PostgreSQL Docker walk-forward backtesting model calibration experimental design Claude Code


背景 | Background

  • 量化研究 — WorldQuant BRAIN Research Consultant(2026.08 起,有薪顧問合約,Gold 等級):交付 38 個通過品質與穩健性檢查的全球股票訊號(樣本內 Sharpe 中位數 1.63)。 Research Consultant at WorldQuant BRAIN (since Aug 2026, Gold level): 38 global equity signals passing quality and robustness checks (median in-sample Sharpe 1.63).
  • 業界 — 優照護(UCarer)R&D 專案經理(2021.03–08):整合性失智症延緩方案,協調認知訓練、睡眠偵測、營養產品三個外部合作團隊,參與雲端平台與 API 規劃。 R&D Project Manager at UCarer (Mar–Aug 2021): a three-partner dementia-delay program, including cloud platform and API planning.
  • 研究 — 台大 & 中研院跨領域神經科學博士候選人,預計 2027 年 3 月畢業。兩篇第一作者論文(Journal of Psychiatry and Neuroscience 2023、Biological Psychology 2021)。 PhD candidate, NTU & Academia Sinica, expected Mar 2027. Two first-author publications.
  • 國際 — 2024–2025 以教育部公費在美國波士頓擔任 Brandeis University 訪問學者近一年,協調台美跨站點神經影像資料平台,日常以英文與美國團隊協作。同期擔任波士頓台灣生技協會(BTBA)年度研討會募款共同負責人(募得 US$37K),並在紐約台美生技協會(TAAB)2024 年會擔任籌備委員、以英文主持 keynote(講者為 Graphen 執行長、前 IBM Chief Scientist 林清詠博士)。 Visiting scholar at Brandeis University near Boston (2024–2025, Taiwan MOE-funded), coordinating a Taiwan–U.S. cross-site neuroimaging data platform and working daily in English. Fundraising co-lead for the BTBA Symposium (US$37K raised) and organizing committee member and keynote moderator at the 2024 TAAB Annual Symposium in New York (keynote by Dr. Ching-Yung Lin, CEO of Graphen and former IBM Chief Scientist).

聯絡 | Contact

chunyilee0102@gmail.com · LinkedIn · jamesleecy.github.io

Pinned Loading

  1. fmri-video-feature-analysis fmri-video-feature-analysis Public

    Quantitative comparison of low-level visual, motion, semantic, and object-level information in naturalistic "nature" vs. "urban" video stimuli for fMRI analysis — entropy/energy metrics extracted p…

    Python

  2. line-chat-triage line-chat-triage Public

    LINE 群組健康度 Triage 系統 — 自動分析對話、計算風險指標、產出 PDF 報告

    Python

  3. paper-rag-copilot paper-rag-copilot Public

    A strictly citation-grounded RAG assistant for academic literature libraries, featuring quantifiable hallucination evaluation for trustworthy research assistance.

    Python

  4. semiconductor-quality-ml semiconductor-quality-ml Public

    Anomaly detection and yield prediction across three public manufacturing datasets — a static snapshot (SECOM), in-process time series (PHM 2016 CMP), and a production routing graph (Bosch). Structu…

    Python

  5. precision-weighted-investment-agent precision-weighted-investment-agent Public

    多代理人精度加權投資研究系統 — 以預測處理框架的 precision weighting 合成多個專家 Agent 的台股投資判斷,含走時序回測與校準評估

    Python