Yassawe Kainolda

- Accelerated distributed training of BERT LLM by 16.1%, and CNN models by 14.7%, using gradient pruning approach. - Worked with scaling the training of various LLM models -- BERT, GPT-3. Explored model and data parallelism on multi-GPU systems. - Deployed LLM training on Cloud GPU systems, using AWS and Docker. - Created custom collective communication algorithm by modifying NVIDIA NCCL library, using CUDA, C++

- Improved SOTA EEG brain imagery classification accuracy by 3%, by changing feature extractor embedding to non-Euclidean measure of distance. - My work resulted in a first-author publication on IEEE EMBC Conference

- Developed communication and interconnection systems for IoT networks. - Implemented CSMA (Carrier Sense Multiple Access) algorithm with collision avoidance to achieve multiple access of the common link by IoT devices in the network. - Worked with embedded systems and microcontrollers.

Master's of Science in Electrical Engineering. - Top 1 Science and Technology University in Korea - Thesis: Gradient Pruning to Accelerate Ring-Based All-Reduce in Distributed Deep Learning - Relavant Coursework: NLP, Parallel Computer Architectures, Information Theory

Bachelor of Science in Electrical Engineering - Top 1 University in Kazakhstan - Thesis: Machine Learning-Based Design of Robust Brain-Computer Interfaces