This repository is the complete project catalog behind my GitHub profile. It documents selected systems, research analyses, and applied data projects across healthcare, finance, quantitative research, and decision systems.
For a 30-second overview, visit my GitHub profile. For project scope, methods, evidence, and source repositories, use this page.
| Area | Representative work | Core capabilities |
|---|---|---|
| Healthcare data science | Diabetes risk, ICU ventilation, Parkinson's screening, perinatal risk | Predictive modeling, calibration, fairness, causal inference, clinical data processing |
| Finance & quantitative research | Stock prediction and AI-powered market signals | Financial NLP, BERT, LSTM, time series, signal evaluation |
| Applied analytics & monitoring | Institutional text analytics, SPC anomaly detection | Data standardization, NLP, process monitoring, stakeholder reporting |
| Statistical learning | UCLA coursework and research analyses | Regression, model selection, validation, interpretation |
Built an end-to-end analysis of 253,680 CDC health records spanning supervised learning, subgroup performance evaluation, and propensity score matching.
- Compared logistic regression, random forest, and XGBoost; XGBoost reached ROC-AUC 0.829, PR-AUC 0.428, and recall 0.772
- Evaluated performance across age, sex, education, and income groups
- Improved observed covariate balance to |SMD| < 0.1 for every checked variable after matching
- Estimated a modest association between physical activity and lower diabetes odds while documenting observational-data limitations
Stack: Python · pandas · scikit-learn · XGBoost · statsmodels · causal inference
Developed an educational application that classifies hand-drawn spiral and wave images and explains model attention with Grad-CAM.
- Implemented baseline CNN and MobileNetV2 training workflows in PyTorch
- Added model evaluation, checkpointing, and Grad-CAM visual explanations
- Built a four-page Streamlit interface covering the dataset, performance, predictions, and responsible-use limitations
Stack: Python · PyTorch · CNN · MobileNetV2 · Grad-CAM · Streamlit
Transformed complex, time-stamped ICU data into a reproducible multiclass prediction workflow for invasive mechanical ventilation duration.
- Processed 500,000+ ICU data points and engineered 30+ clinical variables
- Compared random forest and elastic-net multinomial logistic regression
- Reached approximately ROC-AUC 0.80 and Brier score 0.17
- Used calibration, class-specific evaluation, and robustness checks alongside discrimination metrics
Stack: R · SQL · Linux · tidyverse · random forest · elastic net · ggplot2
Team repository · Project site · Supporting research materials
Collaborated in a GDSC AI team to build an end-to-end market research system combining financial text, sentiment signals, technical indicators, and LSTM-based stock prediction. The previously listed “AI-Powered Quantitative Trading Signal Research System” and “Stock Market Analysis & Prediction System” refer to this same project and are consolidated here.
- Contributed earnings-call summarization and BERT fine-tuning
- Helped transform unstructured earnings-call text into structured sentiment features
- Collaborated with teammates integrating news, tweets, earnings calls, VIX, technical indicators, and time-series inputs
- Team-reported project results: MSE 9.73, IRR 12.83%, and win rate 62.38%
Stack: Python · BART · Gemini · BERT · PyTorch · TensorFlow · LSTM · web crawling · financial NLP · time series
Converted inconsistent institutional project records into a reusable analytical dataset and reporting taxonomy.
- Analyzed 4,610 projects from 2014–2024
- Segmented approximately 38,867 Chinese-language tokens with CKIP Tagger
- Organized terms into a 10-category business taxonomy using rules and fuzzy similarity
- Supported reporting and data governance for 10+ stakeholders
Stack: Python · pandas · CKIP Tagger · fuzzy matching · NLP · data visualization
Built a reproducible monitoring pipeline that combines statistical process control with Isolation Forest anomaly detection.
- Calculates I-MR control limits and common SPC signals
- Compares rule-based process signals with model-based anomaly flags
- Includes modular source code, automated outputs, and tests
- Provides a foundation for operational-risk, quality-control, and sensor-monitoring use cases
Stack: Python · statistical process control · Isolation Forest · pytest · Matplotlib
Analyzed healthy, prodromal, and diagnosed groups to identify patterns across non-motor symptoms and clinical indicators. The work emphasized missing-data handling, risk-factor exploration, interpretable comparisons, and responsible clinical interpretation.
Stack: R · SQL · exploratory data analysis · clinical feature analysis
Modeled survival outcomes in U.S. multiple-birth records using plurality-stratified logistic regression.
- Evaluated twins, triplets, and quadruplets with ROC-AUC values of approximately 0.67, 0.73, and 0.79
- Interpreted odds ratios, confidence intervals, and maternal or delivery-related risk factors
- Used stratification and goodness-of-fit diagnostics to avoid relying on a single aggregate model
Stack: SAS · logistic regression · ROC analysis · stratified modeling
Supporting UCLA statistical-learning work covering the foundations behind the larger projects in this portfolio. This repository is presented as technical coursework rather than as a standalone production system.
Across projects, I prioritize:
- Reproducible data pipelines and explicit cohort or target definitions
- Train-test separation and leakage-aware feature engineering
- Metrics matched to the problem, including PR-AUC, calibration, Brier score, and subgroup results
- Interpretable outputs and clear limitations for high-stakes settings
- Translation from model results to operational or research decisions
Languages: Python · R · SQL · SAS
Machine learning: scikit-learn · XGBoost · PyTorch · TensorFlow · CNN · LSTM · BERT · elastic net
Statistics & evaluation: causal inference · propensity score matching · ROC-AUC · PR-AUC · calibration · Brier score · subgroup analysis
Data & delivery: pandas · tidyverse · Linux · Jupyter · Quarto · Streamlit · Tableau · Power BI · Git
GitHub Profile · LinkedIn · Email
Research and educational projects only. Healthcare outputs are not clinical advice, and financial outputs are not investment advice.