Multi-modal Foundation Model

Pretraining

VLM (Vision Language Model)

Generative AI for autonomous driving

1. Worked for the Level-3 Autonomous Driving project as a member of computer vision team

2. Responsible for image-based object detection and lane detection; built the entire 2D object detection pipeline

3. Implemented SSD, Yolov3, and Faster-RCNN object detection models and LaneNet, S-CNN models for lane detection; benchmarked with different backbones like VGG16/VGG19, Darknet53, FPN, ResNet50 and ResNet101

4. Pruned and quantized neural networks and implemented TensorRT runtime pipeline, which saved 40% of inference time

1. Built and modified applications for implementing machine learning algorithms on embedded device

2. Implemented yolo v3 deep learning algorithm for real-time vehicle detection and tracking and license plate recognition

3. Trained the CNN model based on DARKNET neural network framework and reached recall over 90%

4. Created prototype for detecting human key points, gesture and OCR based on TensorFlow framework

1. Conducted event specific multimodal pattern mining, topic modeling and data mining research; extracted city events from Twitter text data using Hidden Markov model, gaussian mixture model and K-means algorithms

2. Implemented Word2vec model and N-gram model for semantic interpretation and refined the corpus on twitter data

3. Used rank constrained regression and fully convolutional networks methods to understand traffic pattern from large-scale web camera videos with low frame rate and low resolution; conducted feature extraction based on OpenCV

1. Designed and managed data analysis pipeline; built SVM and Random Forest models to analyze economic growth

2. Increased accessibility and usability of customer data by redesigning data visualization techniques to include statistical graphs and information graphics and provided data analysis report including data visualization in tableau

3. Built regression model to analyze relationships between different economic indicators and achieving 10% more accurate prediction of performance than previous years