Amended forward shape inference in TensorFlow(e.g. Pad) Utilized backward shape inference to enhance graph partitioning result(200% size in Mask RCNN) Neural network quantization and calibration for specific DLA(including CMSIS-NN & Andes NN library backend, used in Tiny ONNC ) Customized neural network inference engine with MLIR for specific DLA simulation Achieved 99.9+% accuracy in int8 mode of specific DLA for DLRM model in a tight schedule Built deep learning model calibrator with MLIR Studied state-of-the-art deep learning model quantization/calibration methods and implement them in C++

ai/ ). Developed just in time compiler for image accelerator, used in freeRTOS and Embedded Linux environment. Pre- and post-silicon verification(NPU and image accelerator). Debug mismatch between CSim and RTL or performance issue with firmware, CSim, and HW team. Survey advanced techniques, compiler: TVM, MLIR, runtime: TFRT, IREE, model/hardware co-design: network architecture search. Recruit and lead interns to complete specific research work. Fongfu．Software Engineering Manager．2016 / 9 ~ 2018 / 6 Build up a software team (~5 engineer) to deploy services on iOS/Android and Asus Zenbo