MANU
MATHEW
JISS

Endpoint-Sensitive Causal Audits of Mixture-of-Experts Safety

Audited one-step routing interventions in OLMoE and Qwen3-MoE, separating a changed next token from attention-cache effects and comparing answer-prefix grades with complete-answer grades. In the reserved misinformation study, routing changes often rewrote later text but produced no complete-answer compliance-label changes. Anonymous submission under review at ICLR 2027.

LIMFADD: LLM-Enabled Instagram Multi-Class Fake Account Detection Spring 2024

Developed LIMFADD, an LLM-generated dataset for Instagram fake account detection enabling classification into real, spam, scam, and bot categories with 97% accuracy. Designed the data-collection and XAI-based validation pipeline; paper published at IEEE ISTAS 2025 with TechRxiv preprint.

International Symposium on Technology and Society, Santa Clara University, CA

Dr. Tapadhir Das, University of the Pacific

A Three Phase Pipeline for Sentiment, Toxicity, and Account Authenticity on Social Media Comments

Written as a Graduate Research Assistant under Dr. Tapadhir Das. Developed a three-phase AI pipeline combining sentiment analysis, toxicity detection, and account authenticity classification using XGBoost and LIME, achieving 96.9% accuracy. Designed an end-to-end system linking comment sentiment with account-level risk profiling for bot, scam, spam, and real account detection. Paper published at IEEE SMC 2026.

IEEE International Conference on Systems, Man, and Cybernetics, Bellevue, WA

Dr. Tapadhir Das, University of the Pacific

Sentiment Trajectory Analysis in Software Update Reddit Discussions

Large-scale trajectory research on 324+ Reddit discussions in the ReleaseTrain.io ecosystem. Built a Python pipeline using VADER, REST cohort APIs, and custom reliability metrics comparing author versus community sentiment, validated via human labeling and confusion matrices.

Dr. Solomon Berhe, University of the Pacific

AI-Driven Perception Stack for F1TENTH Autonomous Racing

Hands-on autonomy track with Dr. Dongbin Lee on NVIDIA Jetson Orin Nano and ROS 2 Humble: fused LiDAR and ZED stereo streams, waypoint following, reactive obstacle resets, CUDA-aware Docker deployments, and regression-grade ROS bags. Update: layered in SignSight, a template-guided vision-language framework blending CLIP-style semantic cues with geometric sign templates so traffic-sign hypotheses stay stable at racing speeds. The integrated Jetson-centric study is undergoing peer review for a CVPR 2026 submission on OpenReview while F1TENTH experimentation continues for classroom demos.

Dr. Dongbin Lee, University of the Pacific

AIMoodDiary: Hybrid Emotion Inference for Reflective Journaling and Mood Analytics

Built an end-to-end affective journaling system integrating transformer-based emotion classification, constrained LLM diary generation, and longitudinal mood analytics. A deployed RoBERTa 3-class model reaches 74.6% accuracy and 74.4% macro-F1 on a balanced GoEmotions split (n = 5,421). The diary pipeline combines user keywords and generated text via 72/28 weighted fusion with lexical calibration, evaluated through a five-configuration ablation and fusion-weight sensitivity analysis. Deployed as a live, non-clinical reflective wellness platform; submitted to IEEE ICMLA 2026.

Explainable Large Language Models for Cardiovascular Disease Detection: Opportunities, Challenges, and Ethical Considerations

Single-author white paper using a mixed-method systematic review of 22 peer-reviewed articles plus thematic analysis to examine how Explainable AI (XAI) can address the black-box nature of LLMs in cardiovascular disease detection. Surveys SHAP- and LIME-based interpretability across EHR, ECG, and multimodal clinical pipelines, and synthesizes the technical, ethical, and adoption challenges for transparent, trustworthy clinical decision support.

University of the Pacific, Department of Computer Science