I build end-to-end data systems — from LLM pipelines to analytics dashboards.
Currently pursuing a Master of Data Science at UBC.
- 🔍 Focus: LLMs / NLP · Data Pipelines · ML Experimentation · BI
- 🎓 Background: Economics + Data Science (AI & ML)
- 💼 Open to: Vancouver · Toronto · Montréal · Remote / Relocation
- 📬 Reach me: yinwentsai@gmail.com · LinkedIn
Languages: Python · SQL · R
ML / NLP: PyTorch · scikit-learn · Transformers · LangChain
Cloud & Infra: GCP (BigQuery, Vertex AI) · Azure OpenAI · Docker
BI & Analytics: Power BI (DAX) · Tableau
Automation: Power Automate · n8n
🔹 Mental Health LLM Classification (repo coming soon)
Multi-model pipeline comparing LLM prompt engineering vs traditional ML for 7-class mental-health text classification on 53K social media posts. Identified and corrected data leakage in a widely-referenced Kaggle baseline.
SVM macro-F1 0.735 · Gemini 2.5 Pro F1 0.605
Python LLMs scikit-learn Gemini Kimi-K2 Prompt Engineering
🔹 DQN Quantitative Trading Pipeline · repo
MLOps pipeline combining a DQN agent and CNN Q-network for trading signal generation on Taiwan Weighted Index (^TWII, 2016–2024). Includes Airflow scheduling, MLflow experiment tracking, and a Streamlit + LLM report UI. Extended from a group project with full pipeline rebuild.
Sharpe 0.504 → 1.061 after hyperparameter tuning · compared against 5 baselines
Python TensorFlow DQN MLflow Airflow Streamlit
🔹 ICU Extubation Prediction · repo
Clinical ML model predicting extubation readiness from MIMIC-IV time-series EHR data, with SHAP-based interpretability and subgroup analysis. Currently reworking with corrected methodology and confidence intervals.
AUROC 0.822 · Accuracy 80.66% (Neural Network, best of 5 classifiers)
Python SQL scikit-learn SHAP MIMIC-IV
🔹 MLP vs DOE for A/B Testing · repo
Python reimplementation of a CUHK SRPP research paper (originally MATLAB). Identified and fixed a bug causing PCS to plateau at 0.5, and added a head-to-head budget comparison between MLP and factorial designs — missing from the original paper.
PCS convergence verified · bug fix + equal-budget comparison added
Python NumPy DOE Experimental Design Research