About Me
I am a Machine Learning Engineer at Google, where I work on Gemini Factuality. Specifically, my work focuses on improving grounding in Gemini so it better adheres to context documents when answering questions in an Open-Book QA/RAG style. Before my time working on Gemini, I worked on high-performance LLM training on Cloud TPUs and GPUs, building Google's popular MaxText training framework and leading Google's submissions to MLPerf Training for Dense LLMs.
Before my time at Google, I was an MSCS graduate student at Stanford University, where I specialized in Deep Learning and NLP. At Stanford, I had the chance to perform exciting research projects under the guidance of Christopher Manning, Christopher Potts, and more on machine translation, AI for Healthcare, and beyond. I also had the privilege of teaching Deep Learning under Andrew Ng and mentoring numerous amazing projects. I also completed my undergraduate studies at UC Berkeley in Computer Science, where I learned under the tutelage of amazing professors like John DeNero, Alexander Paulin, and Satish Rao.
I also spent five years working at Salesforce and Salesforce Research, both as a Senior Software Engineer and Research Engineer. I was one of the key engineers for the launch of Customer 360 Data Manager, implemented numerous enterprise-scale microservices, and trained and tested LLM-based multilingual neural machine translation models to rapidly localize Salesforce help texts.
In my free time, you can catch me voraciously consuming podcasts, listening to audiobooks, reading up on the latest Hollywood trade news, and at the gym.
Selected Works
Gemini 2.5 Technical Report
A key contributor to the Gemini 2.5 model family, with a focus on enhancing Factuality and Grounding. My work involved contributions to the model's training, comprehensive evaluation, and the development of the FACTS benchmark.
MLPerf 4.1 Training Submission
Led Google's MLPerf 4.1 LLM Training Submission, overseeing GPT3-175B training on up to 6144 v5p TPUs using Jax and MaxText; improved E2E training step time over Google's 4.0 submission by up to 8% through better overhead optimization, parallelism sweeps, and beyond.
Official JAX for PyTorch Users Tutorial
Developed and published a tutorial + Kaggle Notebook for Google Cloud Blog that introduces PyTorch users to Jax + Flax NNX by training a simple neural network in both frameworks and connecting core concepts, in close collobration with Jax and Deepmind.
How Low Can You Go? A Case Study in Extremely Low-Resource NMT
Case-study, conducted under the guidance of Christopher Manning, on adapting neural machine translation to extremely low-resource language settings for language preservation, focusing on Cherokee to English translation.
SHAZAM: The Effects of Pre-Training and Fine-Tuning on Cross-Domain Sentiment Analysis With ELECTRA
Thoroughly experimented on the benefits and drawbacks of additional pre-training and fine-tuning of LLM encoders for cross-domain sentiment analysis.
Customer 360 Data Manager
Lead engineer on Data Federation Service for Customer 360 Data Manager from 2018-2022, the marquee product launched at Dreamforce 2018 and 2019. Created the microservices and transformation libraries that powered the exchange of data across different data systems and schemas.
Content Recs
A few books, movies, podcasts, and more that I've been enjoying recently.
Books
Project Hail Mary
A lone astronaut, the sole survivor of a last-chance mission to save the Earth, must unravel a scientific mystery to prevent humanity's extinction.
The Planets
A collection of essays exploring the history, mythology, and science of each celestial body in our solar system.