Mohammad Donyavi

Hi, I’m Mohammad — A Passionate and versatile AI researcher with a strong foundation in deep learning, LLMs, and distributed ML systems. Experienced in optimizing model performance and building scalable data/ML pipelines. Adept at bridging academic research and practical deployment in NLP systems..

My Journey

My academic and technical journey began with a foundation in engineering problem-solving during my undergraduate studies in Petroleum Engineering at Sharif University of Technology. Fascinated by data-driven approaches, I focused my bachelor thesis on applying machine learning to analyze wettability changes in smart water injection — an early sign of my growing interest in AI and computational methods.
Driven by a passion to bridge engineering challenges with intelligent systems, I transitioned into the world of computer engineering for my graduate studies. I was accepted into the prestigious University of Tehran, where I began my M.Sc. in Computer Engineering with a concentration in Computer Architecture and AI systems. During this time, I immersed myself in advanced coursework spanning deep learning, distributed systems, and statistical inference, all of which laid the groundwork for high-performance AI research.
My master's thesis, titled “Accelerating Deep Neural Networks Training for Large Language Models,” focuses on optimizing large-scale model training through multi-constrained computational graph partitioning and distributed training techniques. I worked extensively with frameworks like PyTorch Distributed, DeepSpeed, and FSDP, reducing communication overhead by 23% and enhancing GPU efficiency by up to 33%.
In parallel, I gained hands-on experience with cutting-edge large language models (LLMs) such as LLaMA 3.1, Phi-3, and Gemma-2, fine-tuning them on curated datasets and evaluating them using rigorous NLP benchmarks like MMLU and COMET. I also contributed to data preprocessing pipelines for low-resource Persian datasets, improving classification performance by 39%, and explored Retrieval-Augmented Generation (RAG) for enhancing text generation systems.
My commitment to expanding my global academic perspective led me to Mälardalen University in Sweden as an Erasmus+ exchange student, where I studied foundation models, intelligent systems, and advanced algorithms — enriching my understanding of cross-disciplinary AI applications.

Experiences

Research Assistant at NLP Lab [University of Tehran] - (Sep. 2021 – Aug. 2025)

  • Developed a multi-constrained computational graph partitioning approach to parallelizing a deep neural network model (LLM) across GPUs, resulting in an average 23% reduction in communication overhead and up to 33% improvement in memory and compute efficiency. (PyTorch Distributed, METIS, PEFT, DeepSpeed, FSDP)
  • Evaluated fine-tuned large language models, including Llama 3.1 (8B–13B), Phi-3 (3.8B), and Gemma-2 (2B), achieving an average 16.3% improvement in perplexity through rigorous testing on benchmarks such as MMLU, BeLU, and COMET. (LM Evaluation Harness)
  • Preprocessed and cleaned low-resource Persian datasets (e.g., Cultura-X, Alpaca), resulting in a 39% increase in classification accuracy on downstream NLP tasks. (Fairseq)
  • Employed Retrieval-Augmented Generation (RAG) to enhance model performance by integrating relevant external information into generation tasks (Hugging Face Transformer)

    • Django Web Developer Intern [Mapsa] - (May. 2020 – Sep. 2020)

  • Collaborated with a team of 15 to design and develop back-end sites. Utilized Django, Rest Framework, HTML, CSS, JS, SQL, No-SQL, and Postgres to create efficient and user-friendly websites. Contributed to the design and development of the back-end architecture, ensuring optimal site performance and functionality.
  • Github Link

    • Teaching Experience

  • Neural Networks and Deep Learning - Dr. Kalhor (Fall 2024)
  • Machine Learning - Dr. Babak N Araabi and Dr. Mohammadreza Abolghasemi Dehaqani (Fall 2024)
  • Natural Language Processing - Dr. Dr. Hesham Feili (Fall 2024)
  • Microprocessor Laboratory - Dr. Sheikhaee (Spring 2024 - Spring 2025)
  • Advanced Computer Mathematics - Dr. Yazdani (Fall 2023 - Fall 2024)

    • Skills

    Honors

    Example: Ranked 25th among 40,000 participants in Iran’s national Computer Engineering entrance exam — top 0.06%

    Get In Touch

    Whether you're hiring, collaborating, or just curious — feel free to reach out. I'm always open to exciting opportunities in the world of data and AI.

    • Address

      Tehran.
      University of Tehran,
      College of Engineering,
      Electrical and Computer Engineering (ECE) Department

    • Phone

      +98 937 204 9946

    • Email

      MohammadDonyavi1999@gmail.com