Delesley Hutchins
Principal Software Engineer, Generative AI
Campbell, California, United States
Summary
🤩
Rockstar🎓
Top SchoolDelesley Hutchins is a Principal Software Engineer with 14+ years of experience at the intersection of compilers, programming languages, and large language models, currently building generative world models on NVIDIA’s Cosmos team. He holds a PhD in Informatics from the University of Edinburgh and has shipped production compiler features (Clang thread-safety analysis) as well as pioneering research systems at Google and DeepMind, including TensorFlow Fold, the Memorizing Transformer, Block-Recurrent Transformer, and Melodi. His work enabled 100x speedups for neural models with divergent control flow on TPUs and produced one of the earliest truly long-context transformers handling 250K tokens on a single TPU core. Author of the meliad open-source JAX/Flax transformer library and contributor to LLVM/Clang and TensorFlow projects, he combines deep systems-level expertise (XLA, CUDA, TPUs) with practical ML architecture design. Uncommonly for an LLM researcher, he retains strong roots in static analysis, type systems, and DSLs, and has unpublished work on neural tokenization and latent-space language modeling that reflect a long-term interest in making models editable and memory-efficient.
14 years of coding experience
31 years of employment as a software developer
PhD, Informatics, PhD, Informatics at The University of EdinburghB.S., Computer Science, B.S., Computer Science at The University of Texas at Austin
Github Skills (13)
compiler10
compiler-design10
compiler-compiler10
c-language10
deep-learning10
tensorflow10
cprogramming-language10
python10
concurrency10
benchmark9
benchmarking9
security8
coding-style8
Programming languages (3)
Github contributions (5)
tensorflow/fold
Feb 2017 - Oct 2017
Deep learning with dynamic computation graphs in TensorFlow
Role in this project:
Back-end Developer
Back-end DeveloperContributions:17 commits, 4 PRs, 14 pushes in 8 months
Contributions summary:Delesley's commits focused on updating benchmark code used in a research paper, specifically moving the code to a new directory structure. The user also made minor bug fixes to the benchmark code, which included updating file imports and documentation. These changes indicate a focus on maintaining and improving existing functionality and aligning with the project's goals of deep learning.
computationcomputation-graphsdeep-learningmachine-learninggraphs
llvm-mirror/clang
Sep 2011 - Sep 2015
Mirror kept for legacy. Moved to https://github.com/llvm/llvm-project
Role in this project:Back-end Developer & Security Engineer
Back-end Developer & Security EngineerContributions:158 commits in 4 years 1 month
Contributions summary:Delesley's commits primarily focus on implementing and refining thread safety analysis within the Clang compiler. Their contributions include adding new warning flags, refactoring existing code to enhance maintainability, and introducing new attributes for fine-grained control over lock behavior. The work also includes fixing issues related to the handling of smart pointers, array accesses, and potential deadlocks, significantly improving the robustness of the analysis. The user's work demonstrates a deep understanding of compiler internals and concurrent programming principles.
keptwindowsllvmcc-plus-plus
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.