Rui Wen

Close Automation

- Built a REST API with FastAPI and open-source LLM Llama 2 7B Chat to generate personalized writing feedback analyzing grammar, structure, and content quality for 300+ students, improving feedback turnaround time by 75% - Experimented with Tensordock servers, Google Colab, AWS SageMaker to achieve a cost-efficient solution for reliable model-generated reports, reducing the overall cost by 40% compared to the initial Google Cloud solution - Fine-tuned Llama 2 7B Chat with 800+ human-graded reports using supervised fine-tuning and QLoRA, achieving 55% improvement in grading accuracy compared to GPT-4 baseline

- Developed a full-stack web application game connecting 500+ users through real-life and virtual walk videos - Built backend infrastructure using Node.js, TypeScript, JWT authentication, and MySQL on AWS RDS - Implemented lazy loading for offscreen images and non-critical React components, optimizing page load time by 30% - Instituted CI/CD workflows with GitHub Actions to automate deployment, reducing release time by 40%

- Automated catering order management by building a comprehensive order tracking system with Express.js and Wix APIs - Developed dynamic dashboards with ingredient tracking, top customer insights, time-based filtering, and CSV export by integrating with Wix Data API, enabling real-time visibility into key business metrics for stakeholders - Improved the business’ efficiency in order processing and food preparation by 90%, saving 4 hours of manual work per week

- Developed a Spotify song recommendation website with Flask and Spotify API to queue songs from a user’s playlist based on the current mood that the user inputs - Tested and compared different machine learning algorithms for the recommendation system - Optimized the model with K-Nearest Neighbors algorithm to identify the mood of a song with 70% accuracy - Won the 2022 IEEE Northwestern Best Design Award among 23 competing teams

- Investigated solutions to separate mixed X-ray Fluorescence image of an overpainted oil painting for art analysis and conservation - Implemented an autoencoder with convolutional neural network and PyTorch to extract features from a hyperspectral imaging dataset; achieved 78% accuracy in image reconstruction with the autoencoder - Utilized multiple remote computing resources such as Northwestern University's GPU system QUEST to handle big amount of data - Awarded summer research grant by Partnerships for International Research and Education (PIRE)

- Managed the setup for temporary exhibits (Contemporary Brazilian Art Exhibition and “A Sense of Place” Exhibition) - Discussed Dance Lab project and Art Park proposal with local artists - Organized gallery's documents from 2012 to 2019