Profile image
Pin-Ying Wu
Machine Learning Engineer
TSMC, San Jose, CA

About Me

I am a Machine Learning Engineer at TSMC, specializing in LLM and VLM solutions to enhance enterprise decision-making. My expertise spans Generative AI, Computer Vision, and Deep Learning. I independently developed a high-performing internal pricing decision chatbot leveraging RAG and Chain-of-Thought reasoning. Currently, I contribute to a large-scale multimodal agentic system, owning a module that consolidates diverse data sources to enhance decision reliability. I hold an M.S. in Electrical and Computer Engineering from the University of California, San Diego (UCSD), and a B.S. in Electrical Engineering from National Taiwan University (NTU). Beyond industry experience, I have conducted research in 3D Visual Question Answering with Prof. Nuno Vasconcelos at the Statistical Visual Computing Lab (UCSD), and in Audio-Visual Learning with Prof. Yu-Chiang Frank Wang at the Vision and Learning Lab (NTU).

Experience


Machine Learning Engineer

TSMC, San Jose, CA
Apr. 2024 - Present

  • Developed a LLM-powered chatbot leveraging RAG and CoT techniques for internal pricing decisions, and independently delivered a prototype with 92.1% recall, significantly improving decision-making efficiency and accuracy.

  • Contributing to a large-scale organization-wide VLM agentic system, owning the multimodal refinement module that consolidates information from diverse data sources, resolves conflicts, and boosts decision reliability.

  • Collaborating with cross-functional stakeholders to shape project requirements, manage technical risks, and align VLM-based solutions with strategic business goals, leveraging strong technical expertise and clear communication.


    • Graduate Research Intern

    Statistical Visual Computing Lab, University of Californnia San Diego, La Jolla, CA
    Apr. 2023 - Mar. 2024

    Advised by Prof. Nuno Vasconcelos

  • Designed and deployed a data pipeline to collect 100,000 question-answer pairs on safety-related reasoning in 3D scenes, combining GPT-based synthesis for simple cases with MTurk for complex, human-generated examples.


    • Computer Vision Undergraduate Researcher

    Vision and Learning Lab, National Taiwan University, Taiwan
    Sep. 2020 - Jun. 2022

    Advised by Prof. Yu-Chiang Frank Wang

  • Developed an Audio-Visual Transformers model to learn cross-modal contextual features for locating sounding sources in an image, and conducted thorough experimental studies with the MIT-MUSIC dataset.

  • Addressed the fully unsupervised challenge by designing a self-supervised training framework with separate CNNs for visual and audio modalities, incorporating STFT to extract sequential audio features.

    • Selected Projects

    Unsupervised PCB Anomaly Segmentation with Foundational Models Python
    Chih-Hui Ho, SungBal Seo, NaYeon Kim, Pin-Ying Wu, YouSuk Bae, Nuno Vasconcelos
    Electronic Imaging (EI), Intelligent Robotics and Industrial Applications using Computer Vision, Oral, 2024
    Conference paper
    Unveiling the Efficacy of Foundation Models for Depth Estimation Python
    Pin-Ying Wu, Zih-Hao Fu, 2023

    • Explored the potential of large foundation models, including CLIP and Segment Anything, for depth estimation under self-supervised and supervised settings with the NYU-Depth v2 Dataset.

    • Leveraged CLIP's semantic language tokens for initial depth prediction and incorporated parameter-efficient finetuning (PEFT) to study whether these refinement modules can further improve the CLIP predictions.

      • [code][pdf]
    Shallow-PPGNet: A Simple yet Effective Network for Hypertension Detection Python
    Pin-Ying Wu, Ye Sel Lee, Yin Lei, 2023

    • Proposed Shallow-PPGNet, a CNN for detecting hypertension with PPG signals, and achieved over 10% improvement on PPG-BP and MIMIC-II datasets than state-of-the-art approaches.

    • Enhanced the diabetes prediction by transferring the knowledge learned from the hypertension prediction and conducted comprehensive ablations on different prediction models and with various metrics.

      • [pdf]
    An Effective Approach for Arrhythmia Classification using ECG Signals Python
    Pin-Ying Wu, 2023

    • Developed Simple-ECGNet, a light CNN for detecting arrhythmia with ECG signals, and achieved over 10% improvement in sensitivity and precision on MIT-BIH dataset than state-of-the-art approaches.

    • Conducted comprehensive ablations on different prediction models and loss functions with various metrics.

    Face Anti-Spoofing Python
    Pin-Ying Wu, Pei-Ying Lin, Zi-Ting Chou, Chun-Tin Wu, 2020

    • Utilized feature pretraining and sequential modeling techniques to address the face anti-spoofing challenge, resulting in an impressive recognition accuracy of 99.3% on the Oulu-NPU and SiW datasets.

    • Dedicated to brainstorming ideas for model development, encouraging group discussions, setting the project milestones and summarizing the outcomes as a project leader with 3 group members.

      • [code][pdf]
    Visual Domain Adaptation with Adversarial Training Python
    Pin-Ying Wu, 2020

    • Investigated the literature of visual domain adaptation and implemented two popular algorithms, DAAN and ADDA, that adopt adversarial training to align features in different domains.

    • Conducted several domain adaptation experiments with USPS, MNIST-M, and SVHN datasets.

      • [code]
    Face Generation with VAE and GAN Python
    Pin-Ying Wu, 2020

    • Explored and compared the classic generative models, including VAE and GAN.

    • Analyzed the effect of different loss functions and studied the influence of hyper-parameter choices.

      • [code]
    Few-shot Image Classification Python
    Pin-Ying Wu, 2020

    • Implemented Prototypical Network and hallucinated data to classify image with limited training data.

    • Studied different distance metrics and their performance under different settings (i.e. different ways/shots).

      • [code]
    Single-Cycle CPU Implementation Verilog
    Pin-Ying Wu, Si-Ying Chen, 2021

    • A course project collaborating with a teammate in Computer Architecture class using Verilog.

    • Implemented the architecture and functions of single-cycle CPU, supporting RISC-V instructions.

    Controlling DC Motor Position with PID control, lead-lag compensator MATLAB
    Pin-Ying Wu, Hsin-Tzu Sung, 2021

    • A Matlab project collaborating with a teammate in Control System class.

    • Leveraged PID control, lead compensator and lead-lag compensator to implement the desired system and analyzed the property of the system with root locus and Bode plot.

    Source Coding: Turning Signals into Bits MATLAB
    Pin-Ying Wu, 2021

    • Implemented the Huffman tree and the corresponding encoder and decoder, and applied it to convert the resampled and quantized audio waveforms.

    • Analyzed the transmitted and received signals with different noise levels in the communication channel.

    Education

    M.S. in Electrical and Computer Engineering, University of Californnia San Diego, USA
    Sep. 2022 - Mar. 2024
    B.S. in Electrical Engineering, National Taiwan University, Taiwan
    Sep. 2018 - Jun. 2022

    Copyright © Pin-Ying Wu 2025

    Template from Jason Lai