Xiyu Du

- Architected a Retrieval-Augmented Generation (RAG)-based Personal Agent platform on Google Cloud, leveraging Cloud Run for scalable deployment and Vertex AI for embedding to support custom, user-owned knowledge bases. - Designed and optimized high-performance vector search infrastructure using AlloyDB (pgvector) with Private Service Connect (PSC) and VPC peering, ensuring secure access and low-latency semantic retrieval. - Developed blockchain-integrated backend systems for AI model marketplaces, bridging Web3 wallet authentication with machine learning services to support decentralized, tokenized AI applications.

- Collaborated with a research team to developed a text-to-image generator using Variational Autoencoder (VAE) and Conditional VAE (CVAE) models. - Integrated CLIP to align text and image features in a shared latent space for controllable image synthesis. - Trained models on MNIST and CelebA datasets using PyTorch to improve image generation quality. - Conducted hyperparameter tuning, debugging, and expanded experiments on additional datasets to improve generation quality for a reproducible research workflow.

- Analyzed and processed large-scale datasets for text summarization and LLM prompt engineering, utilizing JSON and API-based pipelines for efficient data serialization, storage, and transfer. - Collaborated to develop full-stack interfaces for research presentation using FastAPI, Flask, and OpenAI models. - Helped design and implement a machine learning system to detect and quantify media bias in articles, using NLP techniques to assess text sentiment and ideological leanings across various news sources. - Implemented Retrieval-Augmented Generation (RAG) with embedding models like HuggingFace Transformers and FAISS for vector search, enabling enhanced context retrieval in bias detection tasks. - Contributed to environmental policy research by integrating LLM outputs into equity-focused resource allocation.

- Collaborate to design and develop an online cafeteria system using HTML, JavaScript, and Python, integrating a user-friendly interface for menu display, search functionality, and meal details. - Developed back-end systems to manage daily menus, enabling students to quickly view options, reducing decision time and improving their dining experience. - Implemented data-driven features by incorporating user feedback and dish evaluations through QR codes, enhancing the system's ability to capture real-time dining preferences. - Analyzed food consumption data to predict popular dishes and minimize waste, improving cafeteria operations by reducing food wastage and optimizing supply. - Applied Natural Language Processing (NLP) with LLMs to process and analyze text feedback from students, gaining insights into dietary preferences and making informed adjustments to the meal supply.

- Conducted comprehensive data preprocessing, including data cleaning, transformation, and feature extraction, to ensure high-quality input for machine learning models and enhance predictive accuracy. - Built Python tools for statistical analysis and data mining techniques, extracting meaningful insights from large datasets, and driving data-informed decision-making for engineering projects. - Engaged in structural mechanics simulation and analysis, performing tests on monitoring data and developing simulation models for marine engineering structures. - Identified key demand patterns influenced by various factors, using machine learning models to forecast demand.

- Designed and implemented back-end functions using C++, improving runtime performance and system efficiency. - Analyzed consumer sales data for Coca-Cola Japan, identifying seasonal trends using Excel and Tableau. - Supported database management by designing and optimizing SQL queries to improve data retrieval efficiency. - Maintained data integrity, security, and scalability across ETL pipelines by ensuring reliable data flow.