JunhaoYan jhyan0908@outlook.com · 13048737072 · Shenzhen, China GitHub: https://github.com/jhyan-ux Blog: https://jhyan.xyz/ Email: mailto:jhyan0908@outlook.com arXiv: https://arxiv.org/abs/2110.07376 EDUCATION ========= University of Ottawa (2018.09 - 2022.11) Master of Engineering, Electrical and Computer Engineering University of Shanghai for Science and Technology (2014.09 - 2018.09) Bachelor of Engineering, Optical Information Science and Engineering EXPERIENCE ========== Independent / Entrepreneurship — Founder & AI Product Developer (2024.10 - Present) - Explored B2B AI opportunities in manufacturing after leaving XKL, speaking with factory owners to identify practical AI adoption points and workflow bottlenecks. - Built AICad, an AI-agent product for CNC manufacturing workflows including CAD drawing translation, drawing review, quotation assistance, and version comparison. - Owned backend architecture and implementation end to end while coordinating outsourced frontend work, covering requirements, system design, coding, testing, and demo delivery. - Designed a modular backend using Python, FastAPI, LangGraph, an API layer, business-processing layer, agent library, and persistence layer. Shenzhen Xinkailai Technology (XKL) — Software Development Engineer (2023.03 - 2024.10) - Worked primarily on YMS, an AFM upper-computer software project targeting domestic replacement of atomic force microscope equipment. - Led or independently completed requirement walkthroughs, design, implementation, integration, simulator testing, and clean-room real-machine validation for new features and version quality. - Owned algorithm-module work, developing and integrating multiple data-analysis algorithms and validating them on both simulators and physical equipment. - Implemented hardware adaptation and driver-layer integration in C++ for line-scan and area-scan workflows, improving device interaction stability and performance. - Developed upper-computer service features with Java and Spring Boot, including API design, business logic, and system integration. - Delivered dozens of requirements across 10+ version iterations with no rejected releases, supporting product milestones and customer acceptance. National Research Canada — Computer Vision Research Intern (2020.10 - 2021.10) - Led development of a deep-learning based inspection workflow for aircraft component testing, reducing manual inspection effort through automated image analysis. - Collected RGB camera images, labeled and filtered data with LabelMe, and built a PyQt5 desktop UI for detection results, alerts, live camera view, and before/after image comparison. - Trained and deployed PyTorch and scikit-learn models including DeepLabV3 with a ResNet18 backbone and a random-forest digit detector. - Combined contrastive-learning pretraining, CNN refinement, Adam optimization, and exponential learning-rate scheduling to improve model performance. Lavia Solution — Computer Vision Intern (2020.06 - 2020.09) - Helped build an automatic acne annotation system in the R&D department, focusing on model training for acne recognition. - Annotated image data with LabelMe and trained a Mask R-CNN based detector using Detectron, applying NMS to produce final candidate regions. - Reduced repetitive manual work for skin analysts by automating part of the image-labeling and analysis workflow. PROJECTS ======== Domain Adaptation on Semantic Segmentation with Separate Affine Transformation in Batch Normalization — Research Publication (2019.09 - 2021.09) First Author https://arxiv.org/abs/2110.07376 - Proposed Separate Affine Transformation (SEAT), replacing shared BatchNorm affine parameters with domain-specific affine parameters for source and target domains. - Used DeepLabV3 with a ResNet101 backbone and adversarial training for unsupervised domain adaptation in semantic segmentation. - Evaluated on GTA5 and SYNTHIA as source domains and Cityscapes as the target domain, improving performance by 4.0 mIoU in reported experiments. - Published as arXiv:2110.07376 with co-author Woonsok Lee. DeepLabV3 Defect Detection System — Computer Vision Project (2020.10 - 2021.10) https://jhyan.xyz/work/carrer.html - Built an automated defect/corrosion inspection system for high-temperature aircraft components. - Implemented a two-stage detection pipeline: DeepLabV3 semantic segmentation for coarse corrosion-region detection, followed by multi-layer convolutional edge detection for boundary refinement. - Reached precision and recall of at least 95% according to project records. - Coordinated requirements and data collection with the client, decomposed tasks, assigned work to lab members, and delivered the project on schedule. Embedded Aluminum Measurement System — Embedded Systems Internship (2016.06 - 2016.09) - Developed an embedded measurement system for aluminum products at Huizhou Lichang Aluminum Products Co., Ltd. - Used an STM32 development board to control motor-driven conveyor movement and infrared sensors. - Calculated product length from conveyor speed and sensor timing, reducing manual measurement effort and improving measurement efficiency. Unity Target-Seeking Projectile — Course Project (2020.01 - 2020.04) - Built a Unity simulation environment with vegetation, characters, buildings, and turret behavior. - Implemented movement and action logic in C# and trained a target-seeking projectile with Q-learning. AICad (https://jhyan.xyz/work/carrer.html) AI-agent product for CNC manufacturing workflows, covering CAD drawing translation, drawing review, quotation support, and version comparison. Sole backend owner; delivered an end-to-end demo with Python, FastAPI, and LangGraph. YMS / AFM Upper-Computer Software (https://jhyan.xyz/work/carrer.html) Industrial software project at XKL for an AFM product. Delivered algorithms, C++ driver-layer integration, Java/Spring Boot service features, simulator testing, and clean-room real-machine validation across 10+ releases. Domain Adaptation for Semantic Segmentation (https://arxiv.org/abs/2110.07376) First-author arXiv research proposing Separate Affine Transformation in Batch Normalization for unsupervised domain adaptation in semantic segmentation. SKILLS ====== Backend & Systems: Python, FastAPI, Java, Spring Boot, C++, REST APIs, Modular Architecture, Hardware Integration AI & Computer Vision: PyTorch, OpenCV, scikit-learn, DeepLabV3, ResNet, Mask R-CNN, YOLO, Detectron, Semantic Segmentation, Domain Adaptation AI Agents: LangGraph, Workflow Orchestration, Agentic Backend Design, Manufacturing AI, CAD Workflow Automation Tools & Communication: LabelMe, PyQt5, Unity, Git, English working communication, IELTS 6.5 cv.ha7ch.com/jhyan