Jerry Li
Engineer at Spellbrush
San Francisco, California, United States
Summary
👤
Senior🎓
Top SchoolJerry Li is an engineer with 11 years of experience blending research-grade ML with production software, currently focused on making conversational AI more natural at Spellbrush. He spent five years at DeepMind as a Research Engineer working on projects from reinforcement learning for data center cooling to AlphaFold commercialization, and previously built scalable systems at Google. His open-source work includes contributing ML engineering and deployment features to Twin-GAN for unpaired cross-domain image translation and rigorous test automation for Moshi, a streaming speech-text foundation model. Jerry’s background spans academic research in NLP and vision at Northwestern and hands-on product internships, giving him a rare mix of research rigor and pragmatic engineering. He has a track record of improving model reliability and reproducibility—evident in unit-test driven contributions and disabling non-deterministic GPU kernels to pass CI. Based in San Francisco, he pursues research “for fun,” signaling continuous curiosity beyond his day job.
11 years of coding experience
10 years of employment as a software developer
Bachelor’s Degree Computer Science, Bachelor’s Degree Computer Science at Northwestern UniversityMercersburg Academy
Github Skills (17)
unit-testing10
pytorch10
pytest10
python10
machine-learning10
image-translation10
tensorflow10
trainings10
computer-vision10
modeling10
user-interface9
recurrent-neural-networks9
data-preprocessing9
webui9
javascript7
Programming languages (2)
Github contributions (5)
jerryli27/TwinGAN
Apr 2018 - Jan 2019
Twin-GAN -- Unpaired Cross-Domain Image Translation with Weight-Sharing GANs
Role in this project:
ML Engineer
ML EngineerContributions:24 commits, 1 PR, 19 pushes in 9 months
Contributions summary:Jerry's contributions primarily focused on the development and enhancement of the Twin-GAN model for unpaired cross-domain image translation. They implemented and modified core functionality related to image processing, dataset handling, and model training, including changes to dataset instructions and training scripts. Additionally, the user integrated a web interface with webcam support, demonstrating involvement in both the model and its deployment.
pytorchcycleganweight-sharingdeep-learningtranslation
kyutai-labs/moshi
Oct 2024 - Oct 2024
Moshi is a speech-text foundation model and full-duplex spoken dialogue framework. It uses Mimi, a state-of-the-art streaming neural audio codec.
Role in this project:QA Engineer / Test Automation Engineer
QA Engineer / Test Automation EngineerContributions:1 PR, 1 comment, 2 issues in 2 days
Contributions summary:Jerry primarily focused on adding and modifying unit tests for the `moshi` project, specifically targeting modules related to convolutional and recurrent neural networks, confirming their causal and streaming properties. They wrote comprehensive tests utilizing the pytest framework to validate the behavior of different components, including `SEANetResnetBlock` and `StreamingConv1d`, ensuring the models' expected outputs. The user also disabled cuDNN and made style changes to pass pre-commit checks, demonstrating a focus on code quality and reproducibility.
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.