Yuchi Hsu

• Computer science coursework: Parallel Computing(CS 149), Deep Learning(CS 230)

• Computer science coursework: Data Structure, Algorithms, Web Programming(TA), Computer Programming
• Machine learning coursework: Probability and Statistics, Machine Learning(TA), Applied Deep Learning

• Streamlined translation workflows by fine-tuning LLMs on distributed GPU clusters (8 nodes with 8 A100 GPUs each, using DeepSpeed) for the company's user manual translation project.
• Improved translation quality assessment by implementing neural-network-based evaluation metrics, such as COMET and BLEURT, to ensure more reliable evaluation of translation outputs.

• Improved event extraction accuracy from 86% to nearly 100% by building a prompt evaluation pipeline, enhancing the AI system's data processing.
• Streamlined authorization management for multiple applications by devising a versatile OAuth 2.0 package, improving integration efficiency.
• Expanded AI agent system capabilities by integrating Jira, Outlook, and Confluence, improving cross-platform functionality.

DeepCAT is a startup team that builds an AI-assisted translation web app for patent attorneys (not yet launched). Until now, our prototype has translated 200k+ words for our early-stage users.

• Created a robust patent translation dataset with over 13M tokens, facilitating effective model training.
• Achieved near human performance in patent translation accuracy by fine-tuning LLMs including MarianMT and LLaMA series, improving the BLEU score from 13.66 to 67.85.
• Reduced translation workload by over 50% by developing a rich-text translation editor for patent attorneys.

• Discovered a parameter-manipulation method to reduce the unsafe responses of LLMs from 14.5% to 2.5%, opening the possibility of building responsible LLMs with low cost.
• Co-authored an ACL 2024 paper: "Chat Vector: A Simple Approach to Equip LLMs with Instruction Following and Model Alignment in New Languages"

• Full-stack Development: Next.js, React, TypeScript, JavaScript, HTML/CSS, Node.js, Express.js, MongoDB, Vite, Electron.js, Vue.js
• Machine Learning: PyTorch, Transformers, LangChain, LLaMA, PEFT(LoRA, QLoRA), DeepSpeed, Tensorflow, CUDA
• Others: C/C++, Git, Docker, Verilog, SystemVerilog