Hello, I'm 徐佳茗

Kaylee Xu

I'm currently an M.S. student in Machine Learning at Carnegie Mellon University, graduating in December 2026. Before CMU, I studied Data Science and Finance at NYU Shanghai, where I developed a strong multi-disciplinary foundation in both technical and business domains.

Over the past few years, I've worked on AI-powered production systems at Pinterest, TikTok, and TCL Technology, spanning recommendation systems, multimodal learning, LLMs and Agentic AI systems. At CMU's Machine Learning Department, I conducted independent research under the supervision of Prof. Maarten Sap, working on detecting and mitigating failures in multi-agent systems in social contexts. These experiences have shaped my interests in user modeling, personalization, and cutting-edge AI research. More specifically, I'm dedicated to building AI systems that better understand humans — not only what people say or do, but how their preferences, intentions, and goals evolve over time.

Outside of AI/ML, I'm also interested in stock trading and music composition. If you're working on similar problems—or simply want to chat about AI, markets, or music—I'm always happy to connect.

Experience

Education

Carnegie Mellon University

M.S. in Machine Learning

Kenneth C. Griffin School of Computer Science

Aug 2025 – Dec 2026

New York University Shanghai

B.S. in Data Science & Business and Finance

Summa Cum Laude (Highest Honors) · Double major · Major Honors in Data Science

Aug 2021 – May 2025

Internships

Machine Learning Engineer Intern

LLM-based user journey understanding; sequential behavior modeling and multi-objective notification ranking.

Sequential ModelingMulti-objective LearningLLM Fine-tuningRanking

May – Aug 2026

TikTok

Machine Learning Engineer Intern

Open-world multimodal content understanding for large-scale video recommendation; vision-language fine-tuning on long-tail entities.

Vision-Language ModelsLoRA / Delta-LoRAOpen-world PerceptionRecommender Systems

Aug – Nov 2024

TCL Technology Group

Applied AI Intern

End-to-end LLM-powered speech understanding for smart home devices; 99%+ accuracy with 5× data augmentation. Shipped in commercial products Dec 2024.

LLM Fine-tuningMulti-intent ParsingProduction DeploymentSpeech Control

May – Aug 2024

RedNote (小红书)

Product Engineer Intern

First exposure to AI inside a real product — the turning point toward machine learning.

Summer 2023

Projects

Selected Work

HealthTimeSeriesGym: Agentic AI For MLE Tasks In Healthcare

An evaluation benchmark for AI agents on real-world healthcare MLE tasks over time-series data, emphasizing privacy, robustness, and deployment constraints.

PythonvLLMLangGraphAgentic AI

Research

Multimodal LLM Alignment

Fine-tuned a LLaVA-based multimodal LLM with LoRA on 50K preference-annotated image-text pairs, then trained a DPO alignment pipeline to reduce hallucination — improving MMHalBench score by 12% and outperforming baselines on MMBench, ScienceQA, and LLaVA-Bench by 3–7%.

LLaVADPOLoRAMultimodal AlignmentPyTorch

Independent Study

Consistency Distillation for Diffusion Acceleration

Reproduced Consistency Models and LCM from scratch in PyTorch, exploring modified CFG injection schedules to achieve 8% 2-step FID improvement. Extended to Stable Diffusion and benchmarked against DDPM/DDIM baselines across FID, LPIPS, and CLIP score.

Consistency ModelsStable DiffusionDiffusion Distillation

Course Project

Graph Neural Network Recommendation System

Built an end-to-end LightGCN and NGCF recommendation pipeline with BPR loss on MovieLens-1M, improving Recall@20 by 6% and NDCG@20 by 5% through embedding depth tuning. A popularity-debiased negative sampling strategy added another 2% tail-item recall.

LightGCNGraph Neural NetworksCollaborative Filtering

Course Project

Kaylee Xu

Experience

Projects

Blogs

Life