Ryan Shar
Research Scientist @ Apple
Resume

About

ML Researcher making LLMs practical for real tasks

I have the fortunate of playing with AI models under the guidence of Professor Ameet Talwalkar and Valerie Chen

Current: Research Scientist @ Apple

Email: rshar@cs.cmu.edu

Education

  • MS Machine Learning, 2025

    Carnegie Mellon University

    Overall GPA: 4.0/4.0

    Research: Human-AI Interactions for Programming Tasks

    Relevent Coursework: Graduate Machine learning, ML in practice, convex optimization

  • BSc Honours Computer Science, 2024

    University of British Columbia

    Overall GPA: 4.0/4.0

    Honours Thesis: Improving the weighted error of sparse decision trees

    Relevent Coursework: Advanced machine learning, applied machine learning, NLP,

    intelligent systems, advanced databases, machine learning and data mining, stochastic processes

Papers

  • When Benchmarks Talk: Re-Evaluating Code LLMs with Interactive Feedback (arXiv)

    Jane Pan*, Ryan Shar*, Jacob Pfau, Ameet Talwalkar, He He, Valerie Chen

    ACL, 2025

Experience

  • Research Scientist

    Apple

    2025-Present

  • Graduate Student Researcher

    Carnegie Mellon University

    2024-Present

  • Teaching Assistant

    University of British Columbia

    2019-2024

  • Undergraduate Research Assistant

    University of British Columbia

    2023-2024

  • Firmware Developer

    Motorola Solutions

    2021-2022

Projects

  • User Intent Benchmark

    Carnegie Mellon University

    • Analyzing effects of user ambiguity on LLM performance via simulated human interaction (methods to be published)

    • Implemented novel feedback technique with LLMs using proprietary models (GPT4o, Sonnet) and open source hugging face models (Llama, Gemma, Qwen), achieving 25% higher task correctness

    • Designed and conducted experiments to analyze efficacy of simulated human feedback, showing statistically significant differences in outcome
  • Deterministic & Bernoulli Sampling (pdf)(code)

    University of British Columbia

    • Improved weighted loss of decision trees (GOSDT model) with a novel sampling method using SciPy and numpy

    • Experimentally showed 15% reduction of weighted loss and reduced loss variance compared to models without our sampling method

    • Designed synthetic, weighted datasets with imbalanced and sparse distributions, to represent real data seen in rare diseases and underrepresented populations

  • Class Based VAE (github)

    University of British Columbia

    • Created a novel statistical model for CPSC 440 (Advanced Machine Learning)

    • Achieved improved reconstruction error of scarse labels with lower variance