Preslav Aleksandrov
Education
PhD, Machine Learning
Jan 2024 -- PresentUniversity of Cambridge | Cambridge, UK
- Published research on novel model architectures and training methodology.
- Built high-performance distributed training code for large-scale ML workloads on GPU clusters.
- Authored course materials (slides, notes, examinations) for the Deep Neural Networks course at Cambridge.
MSc, Software Engineering, Distinction
Sept 2021 -- Sept 2022University of Glasgow | Glasgow, UK
- Top of Class award and best final-year project.
- Designed an ML model for efficient cross-device federated learning.
MEng, Engineering, First Class
Sept 2016 -- July 2021University of Glasgow | Glasgow, UK
- Included of Engineering Excellence list 5 years in a row. (Students with over 82% GPA)
- Developed a new optimisation method for highly non-linear systems.
Professional Experience
Research Software Engineer
Sept 2022 -- Jan 2024DeepNano | Glasgow, UK
- Developed finite element method (FEM) code for semiconductor device simulation.
- Designed and trained machine learning models for semiconductor analysis and design, replacing costly numerical simulation steps.
Software Engineer
Jan 2023 -- Jan 2024Semiwise | Glasgow, UK
- Developed machine learning models for semiconductor analysis and design.
- Designed and led a project proposal that secured £500,000 in UKRI funding; managed scope, technical direction, and delivery.
Research Software Engineer
Sept 2021 -- Sept 2022MoFEM | Glasgow, UK
- Developed C++ FEM code for solid mechanics simulation within a large open-source codebase.
Selected Papers
- ABBIE: Autoregressive Block-Based Iterative Encoder for Efficient Sequence Modeling
Developed a novel sequence-modeling architecture using recurrent transformer blocks. Executed full pre-training pipeline to construct the first recurrent transformer scaling laws. - Convolutional Machine Learning Method for Accelerating Non-Equilibrium Green's Function Simulations in Nanosheet Transistors
Applied ML to accelerate quantum transport simulation, combining device physics and deep learning. - The Future of Large Language Model Pre-Training is Federated
Analysed distributed and federated pre-training of LLMs, including the systems-level constraints of large-scale training.
Awards and Honours
- ICASE industrial PhD funding award, sponsored by IMEC (leading European semiconductor and AI research institute).
- ScottishPower Master's Scholarship: selected as top applicant from approximately 30,000 candidates.
Personal Interests
Leather-craft: making small fine leather goods such as cases for glasses and phones.