Jiahui Cao

ROCm and Open source

• Generalized a full wafer defect detection into a shared infra stack by working with division app/algorithm engineers (ingest → GPU- preprocess → transforms → eval) for both serving and training; added job manifests and Ray cluster to reorder millions of defect detection jobs on one common pipeline, demonstrated and reused across multiple KLA divisions. • Designed Spark + SQL–based data-ingestion & feature-engineering flows for defect-image metadata, enabling faster analytics queries and model-training pipelines. • Cultivated proficiency in diagnosing and eliminating performance bottlenecks throughout the pipeline—including disk I/O, CPU compute, GPU compute —while architecting a scalable backend solution for high-throughput processing of millions of die images, incorporating RAM-disk optimizations to minimize I/O latency and maximize throughput (lifting from 200 image/s to 1000 image/s). • Worked in RDMA-enabled RoCE/Infiniband environments for large-scale GPU training clusters, tuning job scheduling and GPU data transfers for throughput and latency optimization. • Implemented and tuned Triton/CUDA kernels for alignment, normalization, and concat operators; additionally optimized a heavily used Triton 2D convolution kernel, improving occupancy and reducing data-layout conversions. Profiled with Nsight Systems/Compute to adjust tiling, num of warps, and memory access patterns, explored op fusion and vectorized/coalesced loads for extra speedups. • Streamlined training and inference workflows for diverse deep-learning architectures in computer-vision applications by refactored the legacy TensorFlow build into a modular Keras-based framework, empowering algorithm engineers to flexibly swap layers and seamlessly integrate custom operations into division-specific graph builders, all within containerized high-performance environments.

• Enhanced key image processing algorithm performance by optimizing kernel functions using Triton and CUDA. • Implemented efficient GPU programming techniques, leading to a 40% reduction in computation time. • Improved data processing speed, resulting in a 35% increase in overall algorithm throughput. • Collaborated with the machine-learning team to achieve a 25% boost in system performance and efficiency.

• Designed a system structure for parallel execution using pytest-xdist. Effectively optimizing the utilization of multiple CPUs. • Developed and maintained a test library for GUI testing. Implemented multithreading in Python Selenium to significantly reduce the execution time from 1 hour to 20 minutes. • Integrated CI/CD processes to automate the testing, integration, and deployment phases of the software development lifecycle. • Implemented Kafka to receive and store testing logs, ensuring centralized management and easy access to test results. • Leveraged React/PHP as the front-end framework to develop a dynamic web page for displaying test results, enhancing usability and accessibility.

• Built RESTful APIs using Node.js and Express.js, ensuring efficient data retrieval and manipulation. • Developed and maintained a full-stack web application using TypeScript, providing end-to-end solutions for clients. • Designed and implemented a robust back-end infrastructure, including the setup and management of MongoDB as the database, and deployed on AWS EC2. • Integrated NLP language analysis techniques to match similar users based on their answers.