SINCE 2023
Senior ML Engineer
- Learning to Learn the Difference
- Parameter-Efficient LLM Adaptation
- LLM Inference on TPU
- Evaluation & RLHF Methodology
Senior ML Engineer · Google · Stanford
George Polovets
Sausalito, California
The engineer who keeps large language models cheap to adapt, evaluate, and serve.
George is a Senior ML Engineer at Google, where he tech-leads LLM inference on vLLM/TPU and evaluation methodology for Gemini. He reached frontier models the long way: a Stanford statistics MS, computer vision at NVIDIA, then language-modeling research at Microsoft and AWS, where his Meta-Learning the Difference (TACL 2022) adapts 1% of a model's parameters to match full fine-tuning. His one original repo still hand-derives a variational Bayes bound for logistic regression.
- 1%of parameters, matching full fine-tuning
- 17.1%kNN-LM perplexity gain, seven domains
- 9 minto adapt one domain, down from 26
- 137GitHub contributions to the vLLM/TPU stack
HAPPENING NOW
STANDING TODAY
HOW THEY THINK4 BELIEFS
Four beliefs run through a decade of language-model work.
THE IMPACT
Ten artifacts across a decade — every one about handling language cheaply.
EVERYTHING ELSE · 4 WORKS
THE TIMELINE
One subject held for a decade — language, and the cheapest way to adapt, evaluate, and serve it.
The whole record, newest first – tap any moment and Aster tells the story.
2023–PRESENT
Making frontier models cheap to serve and honest to measure
At Google since 2023, George tech-leads two mandates: optimizing inference for large open models like DeepSeek-v3 on vLLM/TPU, and building evaluation methodology, data-flywheel workflows, and RLHF post-training for Gemini on Cloud. He is a credited Gemini Team author on the Gemini, 1.5, and 2.5 reports, and logged 137 GitHub contributions to the vLLM TPU stack in the past year.
“Tech lead working on evaluation methodologies of Google's flagship LLMs (e.g. Gemini) on Cloud application areas and researching methods to improve data flywheel workflows and RLHF post training.”
2021–2023
Two ways to adapt a large model cheaply
On AWS AI's Speech Science team, George researched language modeling to power AWS Transcribe. He co-authored two papers attacking one problem — specializing a large LM to a new domain cheaply — from opposite ends: Meta-Learning the Difference (TACL 2022) reparameterized weights, and Adaptation Approaches for Nearest Neighbor Language Models (Findings of ACL 2023) reorganized retrieval memory. He mentored Princeton PhD student Zejiang Hou on the former.
“Researching and developing Language Modeling techniques to power AWS Transcribe.”
2017–2021
Where the language-modeling thread begins
At Microsoft, George did applied research for language modeling and NLP, moving from computer vision into language. The concrete artifact is Long-span language modeling for speech recognition (arXiv 2019), studying whether LSTM and Transformer LMs can carry context across sentence boundaries, co-authored with Microsoft speech researchers Parthasarathy, Gale, Xie Chen, and Chang. Every language-modeling role after this descends from it.
“Applied Research for Language Modeling and NLP.”
2011–2017
A statistician's apprenticeship through analytics, ML, and vision
Before language, George trained as a statistician: valedictorian at Lynbrook High, economics and statistics at Berkeley, then a Stanford Statistics MS (GPA 4.14). Product-analytics internships at Facebook and Yahoo! gave way to his first ML role at Declara — recommendation systems, matrix factorization, word embeddings — then deep-learning computer vision at NVIDIA. His one original repo, VB_Logistic, hand-derives a variational Bayes bound for logistic regression.
“Applied Deep Learning Research for Computer Vision”