Capstone Research Projects

Beyond LLMs: Cognitive-Emotional Trajectories of Human vs. Machine Authorship

Large language models (LLMs) are reshaping text creation, prompting questions about whether human writing trajectories remain distinct. Unlike traditional authorship detection that analyzes isolated texts, we model writing as longitudinal cognitive–emotional trajectories. Using 6,086 human texts and 74,241 LLM generations across academic, blog, and news domains, we capture temporal variability in a 75-dimensional feature space. Humans exhibit richer, more irregular trajectories than LLMs, and our trajectory-based classifier distinguishes them with high accuracy, showing that temporal cognitive–emotional patterns are a robust marker of human authorship.

• Students: YeoJin Jenny Kim, Zhanwei Max Cao 
• Faculty: Shanu Sushmita 
• Subject: Natural language processing 
• Date: December 2025 

Towards Shadow-Invariant Perception: Reflectance Estimation for Robotic Navigation

Robotic navigation is highly sensitive to lighting variation, where shadows and specular highlights distort object appearance and terrain boundaries. We introduce a spectral ratio estimation framework for illuminant-invariant imaging that improves perception robustness under variable lighting. The method builds on classical reflection models while leveraging spectral ratio maps to suppress illumination effects and preserve reflectance. We evaluate performance on raw and benchmark datasets using angular error, shadow detection accuracy, and reproduction error, and further assess downstream impact on object detection and navigation success. Results show that spectral ratio–based representations significantly enhance illumination-invariant perception, offering a practical and physics-grounded solution for reliable robot navigation.

• Student: Yuhan Li 
• Faculty: Bruce Maxwell 
• Subject: Computer vision 
• Date: December 2025 

Drone Ranger

Urban wildlife management in the Bay Area presents growing challenges as animals damage vegetation, attract predators, and create safety risks. This work explores a noninvasive approach using a lightweight autonomous aerial platform to deter wildlife through motion, light, and short acoustic cues. We present a unified obstacle detection and avoidance architecture that integrates perception and control using PX4, ROS2, and Gazebo/Unity, with validation on a Holybro X500 V2 drone. Results demonstrate reliable simulation-to-reality transfer for trajectory planning and obstacle avoidance, establishing a scalable and ecologically sensitive framework for urban wildlife deterrence.

• Students: Zhipeng Ling, Renxiang Yin, Chunzhang Liu, Xiaoman Zou 
• Faculty: Ilmi Yoon 
• Subject: Robotics 
• Date: December 2025

SimPath 2.0: An Emotion-Aware Virtual Patient System for Therapist Training Using Multimodal AI

SimPath 2 is a clinically grounded AI-based patient simulation system designed to support therapist training through realistic, adaptive conversational practice. Building on structured patient personas, DSM-5-aligned clinical knowledge, and persistent interaction memory, SimPath 2 enables trainees to engage in safe, low-stakes therapeutic conversations that evolve in response to their strategies and communication style. The system integrates in-session, rubric-guided feedback to scaffold reflection on key clinical skills such as rapport building, questioning quality, diagnostic reasoning, and empathy. Through this design, SimPath 2 advances human-AI collaboration for professional education by supporting deliberate practice while preserving human judgment and agency.

• Students: Minghong Xia, Zixuan Yin 
• Faculty: Akram Bayat
• Subjects: Human-centered computing; artificial intelligence 
• Date: December 2025

Poster will be added when available.

Learning to Play: Game Generation Model with Action and Memory

Recent advances in diffusion models have demonstrated strong robustness in long-sequence generation, motivating their use in interactive game video synthesis. However, existing approaches still struggle with low frame fidelity, inconsistent action-conditioned rendering, and high computational cost. This project develops a lightweight diffusion-forcing framework capable of generating playable Super Mario gameplay directly from data under tight resource constraints. Our method combines a compact 4× VAE with a 32-block DiT diffusion model to preserve essential scene details while maintaining smooth frame-to-frame transitions. Being trained on 522k frames, the model achieves competitive visual quality, including PSNR above 26 dB at short horizons and FID as low as 13.39 at 32-frame rollouts, while sustaining real-time performance at 17 FPS on a single T4 GPU. These results highlight that carefully designed latent diffusion architectures can deliver interactive, action-responsive gameplay without heavy computation. The study provides a concrete case for adapting generative models to interactive environments with limited data, with implications for research, education, and lightweight world-model design.

• Student: Feiyan Zhou 
• Faculty: Tehmina Amjad, Smruthi Mukund
• Industry Partner: Peize Sun
• Subjects: Artificial intelligence; games; machine learning
• Date: December 2025

A Geo-Meta Ensemble Framework for Robust Cross-Well Pore Pressure Prediction

Pore pressure prediction (PPP) is a vital task for ensuring safe drilling in oil and gas operations. Due to the wide application of machine learning (ML) methods, their use is getting the attention of researchers in this domain. The existing methods that apply ML methods for the estimation of pore pressure often exhibit poor cross-well generalization due to domain shift, specifically in regions that have pressure imbalance issues. This study proposes a GeoMeta Ensemble (GME) that attempts to improve these limitations through strategic, well-based data partitioning, physics-informed feature engineering, and a stacking-based meta learner that integrates predictions from diverse architectures, including Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Transformer Model, Deep Feedforward Neural Networks (DFNN), Random Forest, and XGBoost. It employs a Ridge stacking strategy that automatically selects optimal model weights by dynamically leveraging their complementary strengths. This selective ensemble approach combines the use of spatial pattern recognition from CNNs with temporal dependencies of RNNs, attention mechanisms of Transformers, and nonlinear relationships of tree-based models toward robust predictions under domain shift conditions within one model. Processing 267,168 samples from 20 wells across 11 geological regions, the framework sustains an inference latency below 50 ms per prediction, supporting real-time drilling applications. The proposed method achieves R² of 0.9176 on a blind well that has a significant pressure regime mismatch (16.2% underpressure samples in test set vs. 13.8% in training set), thus showing robust generalization under domain shift conditions. The proposed GME outperforms the best individual model (Transformer) by 0.49%, while the transformer model exhibits minimal overfitting (2.7% generalization gap) despite temporal sequence complexities. The study establishes reproducible benchmarks for assessing cross-well generalization.

• Students: Rohan Benjamin Varghese, Pranav Patel 
• Faculty: Tehmina Amjad
• Subjects: Artificial intelligence; machine learning
• Date: December 2025

KN-RAG: A Novel Approach to Retrieval-Augmented Generation with Knowledge Graphs and Self-Contextualization

Large language models (LLMs) have significantly advanced natural language processing (NLP), driving improvements in content generation and question answering. However, these models encounter challenges such as generating non-factual content, high resource demands for updates, and difficulties in effectively processing large inputs. Retrieval-Augmented Generation (RAG) systems offer a solution by integrating factual retrieval to reduce hallucinations. Despite this, standard RAG systems struggle with complex multi-document datasets due to information loss about connections and relationships in the text during vector embedding. To address these limitations, we introduce Knowledge-Nexus RAG (KN-RAG), a novel approach that integrates knowledge graphs with a self-contextualization mechanism within the RAG framework. KN-RAG aims to improve the accuracy, completeness, and diversity of LLM-generated responses, particularly for multi-document question answering.

• Students: Zhixiao Wang, Haoning Wang, Wenqian Xie, Mingyi Qiu 
• Faculty: Ryan Rad 
• Industry partner: Deloitte 
• Subjects: Artificial intelligence; data science; natural language processing 
• Date: Summer 2024