Catherine Lee
Machine Learning Engineer & Software Engineer & Hobbyist Researcher
About
I am passionate about working across the software stack, with expertise in machine learning and distributed systems. In addition to my technical work, I actively engage in research, dedicating my free time to studying the latest advancements in machine learning and contributing to open-source machine learning projects. On any given day, you'll likely find me fitting models on my RTX 3090, eventually resorting to cloud compute--all while wearing a shirt that showcases the Chinchilla Scaling Laws.
Work Experience
Amazon
Machine Learning Engineer
Working on the software engineering and machine learning side for Recommendation Systems at Amazon Prime Video Personalization and Disvoery.
Lamini AI
Machine Learning Engineer
Building the LLM platform for engineers to use customized, private models for their data. Building the training and finetuning abilities underneath, for better performing models compared to general-purpose LLMs.
Software Engineer
YouTube Ads Creative Optimization.
Snapchat
Software Engineer, Machine Learning Intern
Vision Transformers on the Perception team.
Adobe
Machine Learning Intern
Explored multi-arm bandit algorithms for content recommendation (collaborative contextual bandits) such as LinUCB for personalization and used GloVe embeddings for queries.
Amazon
Software Engineer Intern
Built fullstack application in Java using microservices, API Gateway, AWS Lambda, Kinesis, Cloudformation, DynamoDB, S3, CQRS, Event Sourcing, and Dependency Injection. Created a reactive dashboard to monitor the statuses of individual contract workflows with Javascript (SvelteJS).
Education
Stanford University
University of California, San Diego
Skills
Projects
Diffusion LCM with DPO
Latent Consistency Model fine-tuned with Direct Preference Optimization, plus benchmarks comparing sample quality against the base diffusion model at matched step counts.
Llama in Triton
Open-source port of Llama inference to OpenAI Triton kernels. Wrote and tuned the softmax and argmax implementations.
VLM Alignment
Alignment experiments on LLaVA, exploring how preference tuning shifts a vision-language model's responses.
Virtual Try On
Diffusion-based app that renders a garment onto a person's photo from a single reference image.
Place.it
City planning tool for sketching proposed urban changes and visualizing them on a map before they're built.
Bikeable
Scores Boston bike routes by safety, learning from crash and street data to steer riders toward safer paths.
PictRNNary
Multiplayer Pictionary where a recurrent network draws the sketches players race to guess.
Machine Learning Reading Group
A casual reading group I host, working through recent machine learning papers with friends.
Fraud Detection
Click-fraud model predicting whether a mobile ad click leads to a real app download, trained on a heavily imbalanced dataset.
Barcelona Prediction
Models built on Barcelona's open city data to forecast traffic accidents and air pollution levels.