1. events
  2. all
  3. 2025

Jun 6 (Fri) @ 12:00pm: "Modeling Analytic Practices into an AI Agent," Matthew Dupree, ECE PhD Defense

Date and Time

Location: Phelps Hall (PHELPS), Room 1437
Zoom Meeting: https://ucsb.zoom.us/j/81323277734 | pwd=KbpJkkjGUGFtKGY0RvDUQYqAFWZurb.1
Research Area: Computer Engineering
Research Keywords: System Verification, Machine Learning, AI, VLSI Design

Abstract

This defense explores the limitations of current industry approaches to semiconductor test data analytics, particularly with respect to application of generative large language models (LLMs) to the production of AI assistants in this domain. This defense proposes a language-driven view of semiconductor test data analytics. We propose a knowledge graph ontology for representing domain-specific AI assistants, a methodology for extracting domain specific concepts, and a methodology for applying those domain specific concepts to an LLM-powered assistant, including practical demonstration of all three across two assistant designs. This work aims primarily to address the gap in domain understanding for LLMs in the context of low-resource domains of data analytics, as well as the limitation imposed by these domains against misleading outputs and hallucination.

Bio

Matthew Dupree is a PhD Candidate in the University of California, Santa Barbara (UCSB) Electrical and Computer Engineering (ECE) department's "Intelligent Engineering Assistant" (IEA) lab. Matthew joined UCSB from Stanford Online High School in 2016, studied abroad at Lunds Universitet in Fall of 2018, then earned a B.S. in Computer Engineering in 2020. Through the BS/MS program, he then moved under Professor Li-C. Wang's guidance during the uncertain times of COVID-19. In 2021, his undergraduate capstone work "Optically-Guided Multirotor Autonomous Descent and Landing on a Moving Target" earned best undergraduate paper at the International Telemetering Conference. In 2023, his first IEA paper working within the IEA lab, "IEA-Plot: Conducting Wafer-Based Data Analytics Through Chat" earned best paper at the International Test Conference. Since then, he has followed closely the continuing developments in the space of LLMs.

Hosted By: ECE PhD Defense

Submitted By: Matthew Dupree <mdupree@ucsb.edu>