William Tsai
Software Engineer | Chip Designer
- Taipei, Taiwan
- Seattle, WA
Capsule Agent Labs
Powerful agentic AI systems in the making. Coming soon :)
RAG-Based AI Portfolio Assistant
A custom AI assistant leveraging Retrieval-Augmented Generation (RAG) and LLMs to deliver a real-time, interactive portfolio experience integrated into a portfolio site. Features include real-time chat, knowledge/document embedding, rate limiting, session management, error handling, and more.
Thumbnail Recommender for YouTube Videos
A BERT-based multi-label classification model to recommend optimal thumbnail content by analyzing textual metadata to maximize video virality, achieving 90.29% accuracy in multi-label classification.
Data Augmentation for Image Classification using NeRF View-Synthesis
A novel data augmentation pipeline utilizing Neural Radiance Fields (NeRF), improving classification accuracy from 86% (balanced) and 50% (imbalanced) to 91% through synthetic view generation.
Startup and VC Assessment and Matching Platform
A platform connecting startups and VCs through streamlined assessments and an intelligent matching system, enabling fast strategic alignment, fostering collaborations and investment opportunities.
Product Review Generator and Sentiment Classifier
A system designed to generate realistic product reviews using NLP techniques and analyze sentiment accurately through ensemble classification, enabling businesses to simulate customer feedback and gauge customer sentiment effectively.
Investment Management and Strategy Planning System
A web application designed for investment professionals to manage and monitor their investment targets, including setting price alerts, planning trading strategies, sharing trading ideas, and more.
Reconfigurable Radix-2^k×3 Feedforward FFT Architectures
A highly parallel, low-complexity, power-efficient, and reconfigurable feedforward FFT architectures for mixed-radix lengths, ideal for high-throughput mobile applications. A 4G LTE implementation demonstrates up to 33× higher throughput, 7.16× better area efficiency, and 4× lower power compared to previous designs.
A Low-Complexity Mixed-Radix FFT Rotator Architecture
An ultra-low-cost, power-efficient FFT rotator design (a ROM-less and multiplierless architecture). A 128–2048/1536-point implemetation shows 16.75% area reduction and 62.94% power savings with no performance degradation.