Cohere Member of Technical Staff
Making LLM inference faster and more memory-efficient.
Dev Shah
I'm a machine learning engineer based in Toronto. I build inference systems for frontier language models and large-scale RL at Cohere, and explore how agents learn to cooperate, compete, and communicate at Vector Institute.
I studied computer science at the University of Toronto. These days, I build things, run experiments, and write about what I learn.
A few things lately
- workingon LLM inference systems at Cohere
- researchingmulti-agent RL with Zhijing Jin at Vector Institute
- buildingWord Play, a toolkit for text-based multi-agent worlds
- writingabout making lossy KV-cache compression exact in VeriCache
Experience
Vector Institute Research
Multi-agent reinforcement learning with Zhijing Jin.
Shopify Machine Learning Engineer
Pretraining and finetuning generative recommender models.
Vector Institute Research Scientist
Biological foundation models and DNA sequence modeling.
IRCC Full Stack Data Scientist
NLP classification systems and ML workflows on AWS.
Earlier roles
Fallyx Software & ML Engineer
Fall detection with sensor data and deep learning.
University of Toronto Deep Learning Researcher
PointNet models for age classification from 3D bone data.
Robot Vision & Learning Lab Research
Digital twins and robotics simulation in NVIDIA Omniverse.
Interactions LLC Machine Learning Engineer
An LLM-driven avatar for customer support.
Projects
Word Play
Interactive worlds where agents move, communicate, cooperate, compete, and make decisions through natural-language observations.
LLM-driven avatar
A customer-service avatar built with NVIDIA Omniverse and Audio2Face.
Earlier consulting projects
Technical writing
All on Medium- VeriCache: Making Lossy KV Compression ExactInference
- Expressing Neural Networks as Decision TreesInterpretability
- Building GPT from scratchLanguage models
Blog
All postsI also send these out on Substack.