San Francisco Bay Area
- Programming Language: C/C++, Python, JavaScript, HTML, CSS, Verilog - Machine Learning: OpenCV, TensorFlow, Yolo, Object Detection, AI, ML, Deep Learning, Image Processing - Web Application: Bootstrap, jQuery, Django, UI/UX, Full stack, Frontend, Backend - Platform: Linux, Windows, Android, iOS - Misc: Git, Perforce, Makefile, CMake, Electron, MFC, WinForm, EDA, RTL
• Create an internal web application using Django, Bootstrap, and jQuery with Python and JavaScript. This application enables users to compare properties across various configuration files. • Construct an internal desktop application with Electron using JavaScript. This application allows users to analyze the time spent on each hardware phase based on the input log file.
• Design object detection modal aimed to locate and classify the abnormal PV panel. By hyperparameter tuning the modal, final mAP (mean average precision) can reach 0.95. • With the limited training data, we performed data augmentation and preprocessing of 2000 image, which enhance the accuracy by 10%. • At final evaluation, we port our code to NVIDIA Jetson Nano to test its performance. Under the limitation of the RAM and CPU, our algorithm hit the speed of 0.473 sec per image.
• Use C++ to enhance performance while dumping waveform. With the new module compilation technique, we can split our RTL design into smaller sub-designs. It allows multi-tasking, which leads to faster performance. • Design and implement a tool that can create waveform at compile phase. Previously we needed to execute compile and online and offline to generate waveform. With this enhancement, we can directly dump a waveform at compile phase. • Design algorithm to collect multiple select probe command with different depth. For previous use model, we do not allow user to probe individual instance. With this enhancement, we now can provide a more flexible user scenario.
• Used C, C++, and Python to develop a text and image recognition system for Windows, for use in document recognition, text recognition, electronic medical documents, etc. We use state-of-the-art deep learning model to perform object detection and classification. • Achieved scanning image text recognition precision rate of 95%; photographic image text recognition precision rate exceeded 85%; 1D and 2D barcode recognition rate of about 95%. • Developed a smart ID detection and cutting algorithm, using the machine learning object detection model Yolo V3 plus TensorFlow’s CNN to automatically crop the exact position of the ID in a mobile phone camera image; achieved a 95% of IoU (Intersection over Union), with an error rate of less than 1%.