Chen-Han Lin

M.S. in Data Science and Engineering

Coursework: Machine Learning, Deep Learning, Data Science Project, Biostatistics

課程：機器學習、深度學習、資料科學專題、生物統計

B.S. in Mathematics

Coursework: Probability and Statistics, Fourier Analysis, Data Structures and Algorithms

課程：機率與統計、傅立葉分析、資料結構與演算法

Model Enhancement and Implementation of Human Pose Estimation Based on the Mamba State Space Architecture

Integrated Mamba sequence blocks into a Video Pose Transformer for GPU-accelerated monocular 2D-to-3D pose estimation; reduced inference latency and training time through CUDA optimization, parallel data pipelines, and modular integration.

將 Mamba 序列區塊整合至 Video Pose Transformer，用於 GPU 加速的單眼 2D 至 3D 姿態估計；透過 CUDA 優化、平行資料管線與模組化整合，降低推論延遲與訓練時間。

Data Scientist Intern

Systematic Strategy Research: Developed a modular strategy evaluation and statistical validation protocol, testing each rule's excess returns and Sharpe ratio against 10,000-path Monte Carlo random-pick null distributions to separate statistically significant alpha from selection luck.

系統化策略研究：建立模組化策略評估與統計驗證協定，將每條規則的超額報酬與夏普比率，與 10,000 條路徑的 Monte Carlo 隨機選股虛無分布進行比較，以區分具統計顯著性的 Alpha 與選擇結果所造成的偶然性。

Financial Statement Parsing: Built a financial-statement pipeline using rule-based XBRL tree traversal, recovering line items missed by the official SEC company-facts API and validating them against OCR ground truth.

財務報表解析：透過規則式 XBRL 樹狀結構遍歷建立財務報表管線，找回官方 SEC company-facts API 遺漏的會計科目，並以 OCR ground truth 驗證。

Institutional Portfolio Research: Clustered institutional investors from 44 quarters of raw SEC 13F filings using engineered portfolio features—concentration, breadth, and turnover—and direct holdings-weight vectorization; validated cluster quality across silhouette, Calinski-Harabasz, and Davies-Bouldin indices to yield interpretable investor-style segments.

機構投資組合研究：使用集中度、廣度與周轉率等工程化投資組合特徵，以及直接持股權重向量化，對 44 季原始 SEC 13F 申報資料中的機構投資人進行分群；以 silhouette、Calinski-Harabasz 與 Davies-Bouldin 指標驗證分群品質，形成可解釋的投資人風格區隔。

LoRA and Combined Parameter-Efficient Tuning for Large Models

Investigated LoRA, BitFit, Adapter, and Prefix-tuning across vision and NLP tasks. LoRA + Adapter improved vision classification accuracy by +3.6% (0.8775 to 0.9134) with only 1.6% trainable parameters, while LoRA + BitFit improved RoBERTa-base accuracy from 0.931 to 0.945 on SST-2.

研究 LoRA、BitFit、Adapter 與 Prefix-tuning 在視覺及 NLP 任務中的整合。LoRA + Adapter 僅使用 1.6% 可訓練參數，即將視覺分類準確度提升 +3.6%（0.8775 至 0.9134）；LoRA + BitFit 則將 RoBERTa-base 在 SST-2 的準確度由 0.931 提升至 0.945。

Data Science Project: Pima Indian Diabetes Dataset

Owned the full pipeline from raw data to conclusions, building classification pipelines across logistic regression, decision trees, random forests, and XGBoost with preprocessing, domain-specific feature engineering, stratified cross-validation, and grid search. Evaluated the models using Accuracy, ROC-AUC, Precision, Recall, F1, and calibration curves, achieving 83.6%, 91.2%, 84.5%, 88.2%, and 86.3% for the five numerical metrics, respectively.

負責從原始資料到研究結論的完整流程，建立涵蓋邏輯斯迴歸、決策樹、隨機森林與 XGBoost 的分類管線，整合資料前處理、領域特定特徵工程、分層交叉驗證與網格搜尋。以 Accuracy、ROC-AUC、Precision、Recall、F1 與校準曲線評估模型；五項數值指標分別達到 83.6%、91.2%、84.5%、88.2% 與 86.3%。

TSMC IT CareerHack

Built a Docker-isolated refactoring platform with LangChain Deep Agents, FastAPI, React, and MongoDB, automating iterative analysis, refactoring, testing, and quality evaluation across three legacy repositories.

以 LangChain Deep Agents、FastAPI、React 與 MongoDB 建構 Docker 隔離的重構平台，自動化三個舊有程式庫的迭代分析、重構、測試與品質評估。

WorldQuant International Quant Championship

Built a closed-loop alpha research engine that encoded backtest results as graph constraints, pruned redundant hypotheses, and re-ranked signals by marginal portfolio contribution. Evaluated 5,164 alpha expressions, with 1,023 passing all seven platform checks, and helped construct a 52-alpha portfolio that finished 4th in Taiwan.

建構閉環 Alpha 研究引擎，將回測結果編碼為圖約束、刪減冗餘假設，並依投資組合邊際貢獻重新排序訊號。共評估 5,164 個 Alpha 表達式，其中 1,023 個通過平台全部七項檢查，並協助建立由 52 個 Alpha 組成、最終獲得台灣第四名的投資組合。

TSMC IT CareerHack

Fine-tuned LLaVA with LoRA for task-specific visual question answering, supported by targeted data augmentation and noise filtering.

使用 LoRA 微調 LLaVA，以完成特定任務的視覺問答，並搭配針對性的資料增強與雜訊過濾。

TSMC Intelligent Manufacturing Workshop

Formulated the problem as a graph-based job-shop optimization, converting locations to nodes and distances to weighted edges, and applied Integer Linear Programming (ILP) to obtain the optimal schedule.

將問題建模為圖結構的工作車間排程最佳化，把地點轉為節點、距離轉為加權邊，並使用整數線性規劃（ILP）求得最佳排程。

AI CUP 2024 Image Data Generation for UAVs

Built a mask-conditioned image-generation pipeline with targeted augmentation and hyperparameter tuning, evaluated using Fréchet Inception Distance (FID).

建構以遮罩為條件的影像生成管線，搭配針對性資料增強與超參數調整，並以 Fréchet Inception Distance（FID）進行評估。

AI GO Skyrocketed Stocks Forecasting

Built and tuned XGBoost, Random Forest, and TabNet models for stock-movement prediction using feature selection and tabular feature interactions.

運用特徵選擇與表格特徵交互作用，建構並調整 XGBoost、Random Forest 與 TabNet 模型以預測股票走勢。