~/résumé
Bryan Gass
Senior ML research scientist
Developing new recurrent and state-space architectures for efficient long-context learning, from mathematical formulation and JAX/Pallas kernels through pretraining, controlled ablations, and streaming inference.
Broader interests include linear attention, continual learning, learned memory, and embodied agents.
Experience
Systems & Technology Research
- Conduct machine learning research for applied sensing and autonomy systems as a Senior Machine Learning Research Scientist.
working set: Machine Learning · Deep Learning · Python · Research
DEVCOM Army Research Laboratory
- Led end-to-end research on novel input-conditioned state-space and recurrent architectures for subquadratic long-context modeling and associative recall, from mathematical formulation through controlled experiments.
- Translated the architectures into parallel, chunked, and streaming JAX implementations with custom VJPs and recurrent-state caching for training and inference.
- Developed Pallas forward and backward kernels for Triton and Mosaic, achieving up to 26× forward and 41× backward speedups on selected SSM kernels while maintaining numerical parity, gradient equivalence, and streaming/parallel consistency with JAX references.
- Ran 500M–1B-token pretraining and controlled ablations, validated forward/backward training at 1.18B parameters, and built diagnostics for associative recall, gradients and Jacobians, stability, and rare-token behavior.
- Built a modular JAX/NNX research platform spanning distributed GPU/TPU training, evaluation, streaming inference, and deployment, including multimodal APIs and Python/TypeScript SDKs for partner organizations.
- Extended the platform to continual-learning and embodied-agent research through reinforcement-learning objectives, phased training, full-state checkpointing, and robotics environments.
- Presented findings to partner labs and collaborated with academic, industry, and government research teams.
working set: JAX · Flax/NNX · Pallas · Triton · Mosaic · XLA · State Space Models · Continual Learning
Dana-Farber Cancer Institute
- Conducted master’s research on state-space models for resolution-agnostic training and zero-shot upscaling of gigapixel histopathology images.
- Contributed to the development and maintenance of PathML, an open-source computational pathology library.
working set: JAX/Flax · State Space Models · Computational Pathology · PathML
Microsoft Corporation
- Improved multi-GPU training for T5 models on medical documents.
- Investigated state-space, linear-attention, and self-attention methods for extending language-model context.
working set: T5 · Distributed Training · Linear Attention · NLP
Colgate-Palmolive
- Developed machine-learning workflows for supply-chain delay prediction and strategy optimization, supporting cost and emissions reduction and automated risk alerts.
- Built scalable preprocessing and data-management pipelines for large operational datasets.
working set: Python · Machine Learning · Data Pipelines · Supply Chain Optimization
Microsoft Corporation
- Contributed C++ components and developer tooling to the mrecutil library for Microsoft’s automatic speech recognition research and experimentation.
working set: C++ · Automatic Speech Recognition · Developer Tooling
Education
Worcester Polytechnic Institute
Master’s thesis with Dana-Farber: validated JAX/Flax HiPPO operators for resolution-agnostic gigapixel histopathology.
Worcester Polytechnic Institute
Undergraduate research: positive-unlabeled and inductive-logic models for incompletely labeled data.
Cape Cod Community College
Skills
- research
- State Space Models · Recurrent Neural Networks · Linear Attention · Long-Context Modeling · Continual Learning · Generative Models · Reinforcement Learning · Computer Vision · Natural Language Processing · Digital Image Processing
- ml systems
- JAX · Flax/NNX · Pallas · Triton · Mosaic · XLA · PyTorch · TensorFlow · Scikit-Learn · Distributed GPU/TPU Training · Custom VJPs
- engineering
- Python · C/C++ · TypeScript · JavaScript · Java · Git · Linux · Software Engineering · Algorithms · Data Science
- mathematics & data
- Linear Algebra · Differential Equations · Statistics · Probability · Databases · SQL · PL/SQL · Oracle SQL