Yen-Hao Huang
Machine Learning Engineer
Profile
Hello, I am Yen-Hao Huang, a Master's graduate in Computer Science from National Taiwan University. I have two years of experience in machine learning development, working on projects involving chatbot development, RAG systems, and performance optimization tools. Through these projects, I have developed a strong understanding of the end-to-end AI development pipeline, including statistical data analysis, building ETL pipelines, designing models with a research-driven approach, and implementing model training and evaluation in practice. These experiences have strengthened my ability to turn ideas into practical implementations and deliver effective AI solutions.
Experience
NLP Software Engineer
Large-Scale Retrieval
- Designed and implemented a large-scale retrieval system for structured and unstructured data
Research Assistant
Learning-Based RAID Controller Tuning Tool
- Collaborated with HPE to develop a machine learning–based RAID tuning tool, achieving 30% performance gain
- Captured system behaviors using profiler to track cache, memory, disk, and process metrics
- Introduced a rejection mechanism to handle unknown workloads, enhancing model generalization
Education
M.S. in Computer Science
Master's Thesis: Online Tuning of DRAM Controllers Using Continual Reinforcement Learning
- Selected for presentation at the Work-in-Progress Session of the top-tier DAC conference
- Applied reinforcement learning to dynamically tune DRAM parameters for rapidly changing workloads
- Improved efficiency by 35% with a multi-agent continual learning approach
B.S. in Computer Science
Academic Presentation
OT-CRL: Online Tuning of DRAM Controllers Using Continual Reinforcement Learning
- Presenter, Networking Session
Projects
RAG-Based Study Abroad Chatbot
- Developed a chatbot system for FAQs, user experiences, and legal/military consultations
- Integrated LLMs with external data via LangChain
- Optimized chunking and embeddings, reaching 85.5% Precision@5
LLM Sentiment Analysis Tool
- Applied LLM compression (LoRA, quantization, gradient checkpointing) to cut memory usage by 16×
- Fine-tuned model on sentiment analysis task, achieving over 95% accuracy on tens of thousands of reviews
Pothole Detection System
- Developed a detection pipeline for road potholes, achieving mAP@0.5 of 0.82
- Improved accuracy via data augmentation, model selection, and hyperparameter tuning
- Deployed model with FastAPI to enable inference via Web API
Zero-Shot Multimodal Classification System
- Designed a zero-shot learning framework using SigLIP for classifying unseen categories without labeled examples
- Aligned semantic embeddings with vision-text similarity to boost novel class recognition
- Improved ImageNet-1K accuracy from 60% to 73%