Yifei Wang

• Built backend modules using Java, Spring Boot, Redis and Spring Security, supporting role-based access control for different user levels (administrators, inspectors, and enterprise users)

• Addressed the challenge of translating complex government policies (e.g., digital enforcement and electronic evidence requirements) into actionable system specifications for software development

• Designed and contributed to an integrated enforcement platform connecting management, enforcement officers, and enterprise users, resolving interoperability challenges and significantly improving enforcement efficiency through online workflow management, real-time information sharing, and standardized result verification.

• Designed and implemented database schemas and approval workflows to support enforcement planning, review, and execution processes

• Utilized Redis and Node.js to support system performance and asynchronous processing in a distributed architecture

• Collaborated on defining system functionality and technical stack based on real-world enforcement requirements and policy constraints

• Implemented A* search algorithm to solve the 8-tile puzzle with efficient state-space exploration

• Designed heuristic functions (Manhattan, Euclidean) and compared their performance on search efficiency

• Reduced node expansions and improved runtime by selecting optimal heuristics

• Managed priority queue operations (heapq) to maintain O(log n) insertion and retrieval efficiency

• Evaluated performance using metrics such as nodes expanded, solution depth, and maximum queue size

• Using ML to implement automated boundary detection and iterative refinement to accurately crop objects from multi-object and cluttered scenes

• Addressed the challenge of overlapping and ghosting effects in multi-angle radar imaging, where multiple objects and rotations make target identification difficult

• Built a YOLO-based computer vision pipeline to automatically detect object contours from radar cross-section (RCS) images and separate targets from complex backgrounds

• Processed radar return data to isolate single-object images under occlusion and multi-view conditions, enabling more reliable downstream recognition

• Developed contour-based extraction methods beyond standard bounding boxes, allowing precise segmentation of irregular object shapes

• Reduced noise and irrelevant artifacts (e.g., small bright spots and interference) through image preprocessing and edge detection techniques

• Improved recognition quality by generating clean, contour-based object representations for further classification tasks

• Trained CNN models achieving ~80–90% validation accuracy on image classification tasks

• Built and trained deep learning models for computer vision and NLP tasks using PyTorch on datasets with thousands of samples

• Implemented a LeNet-5 CNN for image classification, achieving stable convergence and improved accuracy through hyperparameter tuning

• Trained a character-level GPT model (nanoGPT) on large text corpora (>100K tokens) for sequence generation

• Fine-tuned DistilGPT2 and improved generation quality through decoding strategies (temperature, top-p sampling)