2024 - present
BambooHR
Staff AI/ML Engineer
Leading AI platform work for Ask BambooHR and shared production agent infrastructure: durable runtimes, tracing, observability, evals, and reliable tool use.
Joe Davison
I've been doing machine learning since 2015. Most of the work I'm proudest of has been in NLP: zero-shot classifiers, open-source Transformers tooling at Hugging Face, and later ML for molecular structure and drug discovery at Enveda.
These days I'm sharpening the engineering side of my ML work at my day job, building AI applications and platform pieces close to users. More modeling work to come.
Timeline
A compact version of the path so far.
2023 - 2024
University of Utah
Doctoral studies in computing and AI
Spent a year in doctoral studies with Vivek Srikumar, focusing on NLP, machine learning, and language-model research. Ultimately decided a PhD wasn't the right path for me, so I left.
2021 - 2022
Enveda Biosciences
Senior Machine Learning Scientist
Built Transformer training infrastructure and sequence-to-structure models for molecular identification from noisy MS/MS spectra.
2020 - 2021
Hugging Face
Research Engineer, Science Team
Worked on Transformers and Datasets, zero-shot classification tooling, public models, demos, and research infrastructure.
2018 - 2020
Harvard
MS Data Science
Studied data science at Harvard and interned at IBM Research / MIT-IBM Watson AI Lab, working on cross-population variational autoencoders for isolating latent structure across distributions.
2015 - 2018
BYU
BS Computer Science
Studied computer science at BYU, interned at Microsoft on CNTK, and built early machine-learning projects including MMLSpark NLP components, get_smarties, and devol.
Writing
2020, Blog post
Zero-Shot Learning in Modern NLP
Practical methods for text classification without annotated data; cited in academic literature.
2020, Blog post
REALM: Knowledge and Transformers
Notes on retrieval-augmented language model pretraining.
2020, Hugging Face Docs
Perplexity of fixed-length models
Reference guide for evaluating causal language models with fixed context windows.
2020, Reading group write-up
Attacking Bert: Weight Poisoning Attacks on Pre-Trained Models
Hugging Face reading group discussion and Colab notes on poisoned pretrained model weights.
2021, X/Twitter
Language model prompting and zero/few-shot learning
Thread explaining prompting and zero/few-shot performance across NLP benchmarks.
2018, Medium
Publications
Selected papers. Citation counts are intentionally approximate; see Google Scholar for the changing record.
2019
2019
2020
Transformers: State-of-the-Art Natural Language Processing
EMNLP Demo
Best Demo Award; 14,000+ citations
2021
Datasets: A Community Library for Natural Language Processing
EMNLP Demo
Best Demo Award; 900+ citations
2023
MS2Mol: A Transformer Model for Illuminating Dark Chemical Space from Mass Spectra
ChemRxiv
35+ citations
2022
Multi-scale sinusoidal embeddings enable learning on high resolution mass spectrometry data
ICLR 2023 MLDD Workshop Oral
7 citations
Misc.
Useful projects, demos, and models that do not fit cleanly into papers or writing.
2024, Recognition
AI Utah 100
Named to the inaugural AI Utah 100 list as a Builder.
2025, Patent
Patent US20250189496A1
Patent publication related to mass-spectrometry-based molecular structure prediction.
2020, Open source
Zero-shot classification pipeline
Transformers pipeline and task workflow for classifying text against arbitrary candidate labels.
2021, Open source
Zero-shot distillation
Self-distillation project for training smaller zero-shot classifiers from larger NLI models.
2020, Models
Custom NLI and zero shot models
Public zero-shot and student classifiers on the Hugging Face Hub, including BART, XLM-R, and DistilBERT models.
2020, Demo
Zero-shot classification demo
Interactive Hugging Face Space for classifying text with arbitrary labels.
2017, Open source
DEvol
Early proof of concept for genetic neural architecture search.