Built and deployed a full-stack AI microlearning platform using Next.js, Spring Boot, MongoDB, Kafka, and Anthropic API, enabling users to create personalized learning tracks with LLM-generated syllabi and lessons. Reduced user-facing content-generation latency by decoupling slow LLM and email workflows from HTTP requests using Spring Kafka, 4 core topics, retry handling, and dead-letter queues. Hardened production readiness with Dockerized AWS EC2 deployment, Caddy HTTPS reverse proxy, GitHub Actions CI, SonarCloud analysis, 12 backend tests, and a 200-request authenticated JMeter test with 0 errors and 858.8 ms p95 latency.

Built an end-to-end full-stack AI web application for scoliosis detection from spinal X-ray images using a YOLOv8 object detection model trained on approximately 1,500 labelled medical images, achieving 96% precision in prediction. Deployed a containerized inference service by packaging FastAPI + YOLOv8 with Docker on AWS EC2, enabling authenticated API inference with approximately 0.24 s end-to-end latency per X-ray image. Implemented a scalable image processing pipeline using AWS S3 presigned URLs for direct browser uploads and PostgreSQL via Prisma for metadata storage, enabling secure handling and retrieval of medical scan data.

Developed a modular ML evaluation pipeline within the University of Toronto Machine Intelligence Student Team, achieving a reliability scoring system that assesses autonomous-vehicle perception robustness under adverse conditions. Applied transfer learning with ResNet-18 and ResNet-50 in PyTorch on the WeatherNet dataset of 18,039 images, achieving 89.48% accuracy and 0.897 macro-F1 with ResNet-50. Collaborated with 8 people in an Agile workflow, managing tasks in Jira and contributing to pull-request reviews, design discussions, and weekly standups.

Engineered a full-stack application with 5 RL algorithms and a real-time training dashboard with configurable hyperparameters, enabling users to play blackjack against RL models and sharpen their probability intuition. Benchmarked policy convergence across 1M-episode training runs, demonstrating that learned policies outperform hand-coded basic strategy by 12% in house edge reduction within a constrained action space.

Built a Java desktop trip-planning application using SOLID principles and Clean Architecture. Authored 4,000+ lines of production Java code, leading REST API integration and core accommodation features. Collaborated in a 6-person team, participating in design reviews, GitHub code reviews, and iterative planning.

Developed a virtual tour website for Hoi Ping Chamber of Commerce Secondary School using panolens.js. Integrated panoramic photos taken with Insta360 cameras. Responsible for full-stack development and framework integration under the supervision of the school's IT Panel Head.