Junfei Liu

- Led a team of three in developing end-to-end ML pipelines for automated extraction of cardiac metrics from ultrasound. - Fine-tuned CNN/ViT/SAM for real-time video (over 20 FPS); implemented streaming data smoothing; quantized and deployed models on mobile devices using ONNX and CoreML; developed iOS app and PyTorch migration in Swift. - Designed workflows from frame capture/screening, multi-stage segmentation and measurement prediction, to metrics calculation for diagnosis; implemented with Agentic programming to auto-write/maintain unit tests with context engineering (view-specific instructions + schema hints) and retry + fallback policies, speeding up 2x for dev cycle. - Reducing annotation cost by 73% using strong-to-weak human-in-the-loop distillation through pseudo label generation by a finetuned version of image foundational model MedSAM; implemented a confidence-weighted uncertainty estimation loop, selecting high-disagreement frames for human expert review, maintaining annotation quality while expanding data availability by 16.2x. - Managed weekly Agile cycles with clinicians/CEO; added prompt + tool versioning for agents review.

- Four conference abstracts accepted at The Digestive Disease Week(DDW) 2025 and 2026. - Evaluated MobileNet, Fast-SCNN, and EfficientNet encoders with Unet decoder on liver and abdominal wall ultrasound images, achieving >0.90 Dice coefficient with fastest inference time at 10ms for a single frame. - Proposed and implemented multi-label segmentation for the liver and abdominal walls with shared and separate backbones, achieving 0.938 Dice coefficient with average inference time at 15ms for a single frame.

Working on a math benchmark dataset at Generative Artificial Intelligence Lab(GAIR) of Shanghai Jiaotong University.

Participated as a back-end developer in RocLab responsible for implementing API services under Django framework for web requests of a brand-new course description&scheduler platform with a rating system.

- An internship during which I will construct a machine-learning model by finding and understanding datasets related to specific business problems, performing data wrangling/cleaning/transformation, generating and testing models, and presenting the accomplishment and publishing an article on it. - Involves the use of Agile Framework (Scrum/ Kanban), Github, Python flask, Docker, Postman API, and various machine-learning packages in Python. Focuses on business understanding, model analysis, model deployment (on-prem /cloud), model calibration and monitoring, model result presentation, and team collaboration.

- Teaching assistant for the course Human-AI Interaction - Responsible for guiding group project directions, grading homework and quizzes, and being accessible for various help in coursework, including holding study sessions and office hours. - A course studies recent technology innovations with equity, transparency, and ethics in mind, focusing on the theoretical methods for design and evaluation, which will be implemented in hands-on projects, and introducing special topics, including bias in AI, fairness, future of work.

- Participated in the research sub-project on neural networks and sales models with noise, which aims to train a model that makes predictions as stable as possible in the Covid era, given that existing sales models are inaccurate in today's economy with the pandemic.

- Participated in back-end development using Java. Implemented back-end functions to interact with MySQL and Redis in Controller-Service-Repository structure under Springboot framework in a team of two to improve user experience on the MEDIT Link platform. - Tested and Improved instruction and annotations of an open-source convolutional neural network algorithm, “MeshSegNet,” capable of performing segmentation on scanned human teeth model.

- Honor B.S. in Computer Science, B.A. in Economics, and a minor in Mathematics. - Certificate of Mathematical Modeling in Economics and Political Science.