Jiawei Zhang Email and Phone Number

● Boosted inference speed and saved computational resource usage for 3 GPT-style models by 30%, employing distillation and quantization techniques within PyTorch● Enhanced distributed training performance for two large models by 10.5%, via model profiling and computational graph optimization● Strengthened model reliability by implementing performance metric tracking in Python, with the creation of continuous integration (CI) tests for 3 production machine learning models● Accelerated time to market for new models by 15%, working with other teams to streamline the training and deployment pipeline for multimodal models

● Accelerated Transformer-based models by 12x using novel quantization and pruning methods, contributing them to the Intel Neural Compressor open-source code base in TensorFlow● Improved inference speed by 17x for Diffusion-based models on Nvidia CUDA cores, applying distillation and efficient quantization techniques while maintaining model accuracy● Strengthened reproducibility and collaboration across research teams, adopting model development workflow with MLflow

● Expanded customer base and increased model accuracy by over 11%, by designing and developing 3 multilingual BERT-style LLM using Google JAX● Advanced SOTA models to final product, through modifying model architecture for robustness and fine-tuning with Apache Spark processed data to capture edge cases● Optimized cloud infrastructure and reduced network traffic by 10%, merging two AWS instances and refactoring asynchronous processes and network communications in C++● Ensured the stability and efficiency of production machine learning models with Git and Docker

● Developed the first low-power AI IoT device in the irrigation industry with a real-time embedded operating system, resulting in a 33% reduction in water usage for customers● Reinforced the controllability of the training dataset, managing large dataset using Apache Spark

• Optimized deep learning training time by implementing federated learning with fixed-point operation in Tensorflow, enabling distributed neural network training on 1000 devices• Achieved a 12x speedup on the Nvidia CUDA platform by compressing a neural network model pruning and quantization to compress neural network models

• Designed 5 antennas and extended the products line from 80% to 100% of the commercial frequency range• Wrote more than 12 application papers related to wireless measurement, and awarded employee of the month