Mitchell Naylor

I worked as a technical leader in GitHub Copilot's Applied Science team, leading small teams of applied researchers to execute on high-impact, cross-functional efforts to improve Copilot products. This included:

• Post-training LLMs to improve model performance on high-priority applications, including leading efforts to use reinforcement learning (RL) for agentic tasks — designing reward functions, training pipelines, and evaluation processes
• Designing agentic systems to accomplish complex software engineering tasks, such as reviewing pull requests (see Copilot Code Review), and partnering with engineering and product teams to develop and improve early-stage products
• Conducting A/B tests to quantify model improvements, and contributing to a culture of shipping quickly and rigorously
• Mentoring junior team members through pairing and sharing knowledge of ML best practices, and acting as a key member of the hiring team for applied researchers — designing technical interview material and assessing prospective team members

At Azra AI, I operated as a "full-stack" data scientist, from collaborating with clinical stakeholders to leading research efforts on efficient attention mechanisms for Transformers and model pretraining.

• Led research on efficient architectures for long-context NLP, implementing and adapting emerging methods (including MEGA architecture contribution to Hugging Face Transformers) for clinical NLP
• Designed pretraining and domain-adaptation pipelines, producing language models optimized for deployment on resource-constrained hardware
• Built and deployed clinical NLP models (NER, classification) identifying high-risk findings for 200k+ patients annually, enabling early detection and faster treatment
• Drove technical strategy for ML product development — evaluating cutting-edge research and rapidly prototyping novel architectures for production deployment, and collaborating with clinical stakeholders to ensure model accuracy and reliability met the high standards required for cancer detection workflows

Note: Azra AI spun out as a standalone company in January 2022, previously part of Digital Reasoning

Applied methods from NLP, statistical inference, and operations research to a diverse set of business problems, communicating findings to audiences ranging from highly technical to C-level.

Built predictive models of customer behavior under various pricing scenarios and ran robust statistical analyses in support of the pricing department.