Kuan Zhou
![My image](/static/dea9d3c2e371927579121c42b541c799/41624/d4830a67d9782fa8c34b7493a979add5.jpg)
Contact
- Santa Clara, California
- @kzhoulatte on Twitter
- @kzhoulatte on Linkedin
Passion
Exploring the synergy between science and technology, particularly the exciting field of artificial intelligence.
Specialties
- Programming language: Python, C++, Java, Golang, JavaScript
- AI framework: Pytorch, JAX, TensorFlow
- Distributed system: Torch distributed, DeepSpeed
- ML compiler: MLIR, TVM, LLVM
- MLOps: Docker, gRPC, Kubernetes, Kubeflow, MLFlow, Weights & Biases
- Science: Mathematica, Julia, Matlab
- Others: TeX, SQL, Spark, Hadoop, ORTools, Numba, CUDA
Experience
- Principal Engineer at SambaNova Systems, April 2020 - Present
- Tech lead in containerizing and deploying generative AI models onto Kubernetes platform SambaStudio
- Led a 5+ engineers team to deploy 10+ foundation model based solutions to business customers
- Prototyped the generative AI model deployment pipeline in collaboration with Kubernetes platform team
- Built general and extensive infrastructure for continuous model integration and deployment
- Standardized the model bringup and integration procedure via refactoring ML applications
- Co-designed and co-developed distributed learning infrastructure for extreme large models
- Overlapping gradient synchronization in machine learning(U.S. patent pending with filing date 2/14/2022)
- System of heterogeneous reconfigurable processors for the data parallel execution of applications(U.S. patent pending with filing date 9/9/2022)
- Contributed in core features of SambaNova AI framework
- Designed, implemented and maintained a binary data extractor as bridge between compiler and runtime
- Refactored and upgraded AI framework codebase to support functional programming style dataflow execution
- Implemented various deep learning operators from compiler low level kernels to AI framework end to end
- Optimized performance of deep learning models(HIPNN etc.) based on SambaNova AI framework and dataflow architecture
- Integrated TensorBoard as visualization and accuracy debugger tool into SambaNova AI framework
- Software Engineer - Machine Learning at Petuum, February 2019 - March 2020
- Leveraged OCR engines and deep learning models to process logistic bills automatically with 0.87 accuracy
- Collaborated in implementation of various anomaly detection models for equipment health prediction
- Contributed in machine learning pipeline refactoring and model improvement based on various use cases
- Artificial Intelligence Fellow at Insight Data Science, June 2018 - September 2018
- Architected SketchTML that takes in several hand drawn sketches and produces an interactive HTML website
- Leveraged the framework of pix2code to build a more robust image captioning model with different styles
- Improved BLEU score up to 0.88 through inventive data augmentation methods and weighted loss functions
Education
- PhD in Computational Physics at University of California, Riverside, September 2013 - December 2018
- BSc in Physics, Zhongyao Zhao Applied Physics Elite Class at University of Science and Technology of China, August 2009 - June 2013