Released open-source, “memary” on public (+1K Stars)Created memory modules to remember past personal questions by using Neo4j’s graph storage and Retrieval Augmented Generation of LlamaIndex

• Engineered a React Native mobile app frontend for a health tech startup within UT Austin's startup, using agile methodologies.• Facilitated seamless design iterations by partnering with design teams using Figma and drove our project through a Kanban board.

• Built a MLOps pipeline that preprocesses data, evaluates, trains, and analyzes results with Airflow and versioning datasets and models with Wandb.• Customized stable diffusion inpainting model for wind turbine cracks that are difficult to collect and synthesized 300 images from 12 images. As a result, the YOLOv8-nano instance segmentation performance improved from 0.0427 to 0.718 based on Mask AP@50.

• Developed a Model Experience Web page using NVIDIA Triton Inference Server SDK and BentoML, and optimized AI models by converting PyTorch models to TensorRT, achieving a 1.7x speed enhancement for 16 channel service within Deepstream.• Improved structure restoration, using a custom super-resolution model, made 95 out of 140 hard-to-recognize defects recognizable by adding MS-SSIM loss to EDSR and transplanting Real-ESRGAN’s network to restore defects for LG Display PCB.

• Designed a detection model for webtoon dataset, mixing CIoU and GIoU loss in 3 resized feature maps between ground truth and predicted box and got a SOTA result of COCO AP 76.8.• Designed a classification model to classify nude images in webtoons and analyzed results with GradCam. Improved from 71% accuracy to 97.24% through the rethinking of categorization and CutMix augmentation.

• Collaborated with various teams such as fashion and design was great opportunity to learn cooperation.• Trained DeepLab v3+ to extract clothes from various background• Trained YOLOv3 to recognize 182 types of clothes.• Developed a Resnet-based Siamese Network that recommends which clothes will look good.

• Led a team of 3 engineers to build an end-to-end monitoring system consisting of an ML server, web server, MongoDB, RaspberryPi control board, and android app with 400 CCTVs at 70 farms.• Gathered 245,639 cow mounting motion images – labeled 26139 images by humans and auto-labeled 219500 images by trained object detection model (accuracy: 91% with YOLOv3-spp)• Earned $70,000 by 70 farms subscribing to our platform in 18 months, received $3,000 of investment from Chonnam National… Show more • Led a team of 3 engineers to build an end-to-end monitoring system consisting of an ML server, web server, MongoDB, RaspberryPi control board, and android app with 400 CCTVs at 70 farms.• Gathered 245,639 cow mounting motion images – labeled 26139 images by humans and auto-labeled 219500 images by trained object detection model (accuracy: 91% with YOLOv3-spp)• Earned $70,000 by 70 farms subscribing to our platform in 18 months, received $3,000 of investment from Chonnam National University and $4,000 of winning Capstone Contests. Show less

• Trained model (accuracy: 93% with YOLOv3) to recognize 7 types of vehicles.• Used Deep-sort to track and count them.• Created a dataset of 50k images of 7 types of cars.• Distributed CPU cores by psutil to process more than two videos simultaneously at 32fps.• Created GUIs with PyQt5 and distributed an exe file with pyinstaller and InnoSetup6.