Michael Liang

Master's

Computer Science

CONSTRUCTCONNECT

Data Science Intern

Developed custom post-processing routines for image segmentation models using computer vision methods with OpenCV, Ultralytics YOLO, and PaddlePaddle, significantly improving model detection accuracy and edge alignment while reducing AI output errors by 50 percent. Analyzed model performance on 500 customer images stored in GCS, evaluated by metrics such as IOU, dice, precision, recall, and F1 scores. Deployed models to production through Vertex AI endpoints, delivering updates to 300 enterprise customers via REST API while managing environments with Conda on GCP workbenches. Curated and filtered 150,000 image files to create a high-quality dataset, enhancing model training with convolutional neural networks and vision transformers via Hugging Face.