Liam Li

Research Scientist at Pokee AI

I build intelligent agents for deep research and workflow automation.

Previously, I was an early employee at Determined AI, a Series A startup acquired by HPE, where I led our ML efforts. I have a PhD in Machine Learning from Carnegie Mellon University with over 6,000 citations.

Liam Li

Experience

Research Scientist

Pokee AI 2025 - Present

Building intelligent agents for workflow automation. Developing AI systems that can understand, plan, and execute complex multi-step tasks.

Member of Technical Staff

Fastino AI Feb 2025

Pre-seed startup building task-specific language models.

Distinguished Technologist, ML

HPE / Determined AI 2020 - 2024

Led enterprise LLM solutions including inference services, RAG systems, and finetuning pipelines. Designed DeepSpeed training APIs enabling billion-parameter distributed training. PI on ICML oral paper for cross-modal foundation model finetuning.

LLM Infrastructure RAG Distributed Training DeepSpeed

PhD in Machine Learning

Carnegie Mellon University 2015 - 2020

Advised by Ameet Talwalkar. Thesis on efficient methods for automating machine learning. Developed Hyperband and ASHA algorithms. CMU MLD TA Award. GPA: 4.0/4.0.

AutoML Hyperparameter Optimization Neural Architecture Search

Software Engineering Intern

Google Research 2017

Developed open-source active learning library with 1k+ GitHub stars. Created mixture-based batch active learning methods.

Research Highlights

ORCA

ICML 2023

A general cross-modal fine-tuning framework that extends a single large-scale pretrained model to diverse modalities. Enables efficient transfer learning across vision, language, and other domains.

Paper

Neural Architecture Search

ICLR 2021 (Spotlight) · UAI 2019

Geometry-aware gradient algorithms for NAS achieving near-oracle-optimal performance. Established random search as a strong baseline, improving reproducibility standards in the field.

Paper

ASHA

MLSys 2020

A system for massively parallel hyperparameter tuning that scales linearly with workers. Converges to high-quality configurations in half the time of Google's Vizier with 500 workers. Powers distributed tuning at scale.

Paper

Hyperband

JMLR 2018 · ICLR 2017

A bandit-based approach to hyperparameter optimization that provides over an order-of-magnitude speedup through adaptive resource allocation and early-stopping. Foundational algorithm implemented in all major HP tuning libraries.

Paper