Iz Beltagy
Lead Research Scientist at Ai2
Seattle, Washington, United States
Summary
🤩
Rockstar🎓
Top SchoolIz Beltagy is a Lead Research Scientist and Co-founder/Chief Scientist with 13 years of experience building and shipping state-of-the-art NLP and LLM systems, currently driving the OLMo open LLM initiative at the Allen Institute for AI and co-founding spiffy.ai. He combines deep academic training (PhD, UT Austin) with hands-on engineering—contributions to high-impact open-source projects like Longformer, SciBERT, and SciSpacy show expertise from efficient attention kernels to scientific-domain tokenization and entity linking. Iz moves fluidly between ideation, feasibility studies, and productionization, having implemented optimized GPU kernels and end-to-end training pipelines and Beaker blueprints. Based in Seattle, he balances research leadership with startup product-building, aiming to narrow the gap between open and proprietary LLM research. An under-the-radar strength is his DevOps and data-processing fluency, enabling reproducible large-scale model training alongside novel model design.
14 years of coding experience
16 years of employment as a software developer
Master of Science, Computer Science, Master of Science, Computer Science at Alexandria UniversityPhD, Computer Science, PhD, Computer Science at The University of Texas at Austin
Arabic, English
Github Skills (32)
tvm10
pytorch10
python10
operation10
data-science10
tensorrt10
matrix-multiplication10
gpu-programming10
machine-learning10
ml10
allennlp10
tensorflow10
natural-language-processing10
bert10
cuda10
Programming languages (7)
Github contributions (5)
allenai/longformer
Mar 2020 - May 2021
Longformer: The Long-Document Transformer
Role in this project:
ML Engineer
ML EngineerContributions:2 releases, 9 reviews, 203 commits in 1 year 1 month
Contributions summary:Iz implemented and optimized a custom function `DiagonaledMM` for performing matrix multiplications, likely to accelerate attention mechanisms within the Longformer model. Their work involved integrating and compiling TVM (Tensor Virtual Machine) code for GPU execution, focusing on data types (float16, float32, float64) and device selection (CPU, CUDA). The contributions are critical to the Longformer's performance by efficiently handling long sequences.
pytorchlongformerdeep-learningtransformerlong
allenai/scibert
Jan 2019 - Nov 2019
A BERT model for scientific text.
Role in this project:ML Engineer & DevOps Engineer
ML Engineer & DevOps EngineerContributions:79 commits, 30 PRs, 60 pushes in 10 months
Contributions summary:Iz's commits primarily focus on modifying and creating scripts related to training and data processing for a BERT model used in scientific text analysis. This includes updating and creating scripts to process papers, launch AllenNLP Beaker jobs, and convert text to TFRecords. The commits also show the user experimenting with different models and datasets, indicating a focus on model training and deployment. They are responsible for setting up the Beaker blueprints and configurations for model training.
bertnlpscientific-papers
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.