A curated selection of projects in AI, machine learning, and data science.
Multi-task delivery & satisfaction prediction with MMoE
End-to-end ML pipeline: DuckDB warehouse → feature engineering → Multi-gate Mixture-of-Experts (PyTorch) → ONNX Runtime serving. Predicts delivery time and customer satisfaction simultaneously from 20 engineered features.
End-to-end multi-agent system for semantic code retrieval
Architected a CrewAI + LangChain pipeline with ChromaDB vector search, reducing agent query latency by 40% and indexing 10k+ code embeddings.
Interpretability research on ViT sensitivity to data quality
Used YOLOv5 to generate 15k annotations, applied Grad-CAM + NMF factorization to visualize attention shifts, achieving 18% precision boost.
Multi-task learning with Mixture-of-Experts routing
Resolved cross-task gradient conflicts using MMoE architecture, boosting multi-task prediction accuracy by 4% over Shared Bottom baseline.
Published dataset on visual rhetoric in digital humanities
Co-authored open-access paper in JOHD. Engineered a modular Python pipeline to scrape and clean 3,274 publication records.