Semester of Graduation
Summer 2026
Degree Type
Thesis
Degree Name
Master of Science in Computer Science
Department
Computer Science - College of Computing and Software Engineering
Committee Chair/First Advisor
Dr. Selena He
Second Advisor
Dr. Joy Li
Third Advisor
Dr. Brooke Zhao
Abstract
Job interviews are decisive, yet structured practice stays costly, inconsistent, and unevenly accessible. A systematic review conducted for this work finds existing systems fragmented: most target either technical or behavioral skills, respond weakly to emotion, rarely integrate live coding, and are seldom evaluated reproducibly. To close these gaps, this thesis designs, implements, and evaluates an immersive multimodal mock-interview system realizing the next-generation architecture the review proposed. The platform pairs a Unity WebGL front end and a speech-driven avatar with an ITEM-based scoring backend: a large language model drafts a hidden evaluation, a RoBERTa model reads it, and a sigmoid maps the result to a transparent 0 to 100 score and grade. It adds a live multi-language code editor, emotion and composure analysis, gaze tracking, and a personalized dashboard. Because reproducible evaluation is itself a documented gap, the system is assessed through an automated harness of eight experiments. End-to-end latency averages 12.55 seconds and remains stable across transcript length. The score separates exemplary, marginal, and poor answers at 84.4, 40.3, and 17.1, with a 6.35 percent coefficient of variation. A counterfactual name-swap test finds a small but significant pro-female lean that passes the four-fifths rule at 0.97, and no significant ethnicity bias. An automated judge rates its questions above 90 for relevancy, faithfulness, and precision; the scorer matches human grades on the Mohler corpus at Pearson correlation 0.94. The thesis contributes the first integrated implementation of this architecture and a reproducible, transferable methodology for evaluating AI interviewing systems.
Defense Slides