👋 Hey, thanks for stopping by!
Data Scientist / ML Engineer — LLMs, Python, SQL MSOR @ Columbia University
I turn messy, real-world data into LLM and machine learning systems that move the needle — across recruitment tech, fintech, and analytics. My focus is where LLM research meets real product impact: fine-tuning models, building data pipelines, and shipping systems that outperform far larger baselines at a fraction of the cost.
Highlights
- Fine-tuned a 3B-parameter LLM for person–job matching at Liepin — iterated through 7 versions of Chain-of-Thought prompts and built a full data pipeline reaching 93% precision / 70%+ recall, beating DeepSeek-V3 at a fraction of the cost. Training data curated in SQL, further sharpened with RL reward functions grounded in real user signals.
- Built a Chinese RoBERTa NLP model for customs risk classification — 95%+ accuracy, cutting misclassification by 15% through a domain terminology database and careful statistical analysis.
- At Ant Group, engineered features (crossing, binning, WOE) with Python and SQL to match a GBDT's accuracy in a fully interpretable model.
Currently exploring LLM systems, machine learning, and applied statistics — open to AI/LLM engineering roles as well as data science / analytics / algorithm roles.
Contact: hw3195@columbia.edu · LinkedIn
Unsupervised Image Clustering with JiT Manifold Denoising
Undergraduate thesis project — image clustering pipeline built around a JiT-based manifold denoising model.
