Curriculum Vitae

Python · PyTorch · Scikit-learn · Pandas · Docker · Linux

• Engineered an end-to-end ML pipeline from raw eICU CSV ingestion to model serialization, handling 200k+ patient records with reproducible preprocessing via custom DataLoader classes
• Designed a modular architecture separating data, model, training, and evaluation layers for maintainability and extensibility
• Implemented stratified 5-fold cross-validation harness and automated metric logging (AUROC, AUPRC, F1) across all baselines
• Containerized the full training environment with Docker for environment reproducibility; open-source release in preparation

Python · PyTorch · LSTM · GRU · MLP · Scikit-learn · Pandas · NumPy · Matplotlib · Seaborn

• Designed and trained LSTM, GRU, and vanilla RNN models for sequential data prediction; best LSTM configuration achieved 1.42% MAPE on the held-out test set
• Built a full data preprocessing pipeline: missing-value imputation, min-max normalization, sliding-window feature construction, and train/val/test splitting with no data leakage
• Benchmarked deep sequence models against classical baselines (MLP, Linear Regression, XGBoost) using RMSE, MAE, and MAPE; produced publication-quality visualizations with Matplotlib and Seaborn
• Implemented early stopping, learning-rate scheduling, and gradient clipping to stabilize RNN training on long sequences

Vue.js · Java (Spring Boot) · REST API · SQL · Python · Scikit-learn · Docker

• Built a production-grade full-stack system with a Vue.js SPA frontend calling a Java Spring Boot RESTful backend, with an embedded Python ML microservice exposed via REST for real-time inference
• Integrated trained Scikit-learn classification model into the backend serving layer, enabling on-demand predictions without re-training overhead
• Designed normalized relational schema, implemented JWT-based auth with role-based access control, and deployed the full stack via Docker Compose on a Linux server

Jekyll · Liquid · CSS Custom Properties · JavaScript · GitHub Pages CI/CD

• Designed and implemented a fully custom responsive Bento-grid design system from scratch — zero third-party UI frameworks
• Architected a component-based layout using Jekyll includes, Liquid templating, and CSS custom properties (design tokens) for consistent theming
• Integrated Font Awesome iconography, WebP image optimization, and Web App Manifest for PWA-ready delivery
• Automated deployment via GitHub Pages CI/CD; includes SEO metadata, sitemap, and robots.txt configuration

Deep Learning on Tabular EHR Data · Stroke Mortality Prediction

Designed and implemented a Dual-Tower Transformer (DT-Transformer) for in-hospital stroke mortality prediction using the multicenter eICU Collaborative Research Database. The architecture decouples categorical demographics from numerical vitals into separate encoding towers with Self-Attention, achieving state-of-the-art performance with an AUPRC of 0.6171 (std 0.006) — a 14.4% relative improvement over the strongest neural baseline (NN, AUPRC 0.5394). An Adaptive Runtime Safeguard was integrated to detect out-of-distribution physiological inputs at inference time, ensuring clinical deployment safety.

• Outperformed strongest neural baseline NN (AUPRC 0.5394) and Standard Transformer (0.5279); competitive with XGBoost (0.6467)
• Implemented attention map visualization for clinical interpretability of feature importance
• Paper accepted at CCAI 2026 (The 6th International Conference on Computer Communication and Artificial Intelligence, May 2026)

China Southern Power Grid AI Technology Co., Ltd. · Guangzhou

Python · PyTorch · YOLOv8 · OpenCV · Pandas · NumPy · Linux

• Developed a computer vision pipeline for power inspection robots to automate the recognition of circuit breaker states (open / closed / fault), replacing manual visual inspection on high-voltage transmission lines
• Fine-tuned a YOLOv8 object detection model on a domain-specific dataset of power equipment imagery, achieving reliable detection of switching-state indicators and structural anomalies in varying lighting and weather conditions
• Built and maintained a 60 GB image dataset — covering substations across Guangdong Province — by standardizing annotation formats, deduplicating frames, and writing automated ingestion scripts to feed the training pipeline
• Conducted fault-type analysis via exploratory data analysis (Pandas / Matplotlib), quantifying the distribution of insulator cracking, corona discharge, and loose-connector faults to guide class-balanced sampling strategies
• Contributed to technical bid documentation translating ML algorithm performance benchmarks into procurement specifications for robot deployment contracts
• Outcome: Received written commendation from supervising engineer; earned full score (100/100) on the formal internship evaluation