Join us as the School of STEM master\u2019s degree candidates present their culminating thesis and project work. The schedule is updated throughout the quarter, check back for new defenses.<\/p>\n\n\n\n
View previous quarter schedules<\/a><\/p>\n\n\n\n Select a master’s program to navigate to candidates:<\/strong><\/p>\n\n\n\n Computer Science & Software Engineering<\/strong><\/a><\/p>\n<\/div>\n\n\n\n Cybersecurity Engineering<\/strong><\/a><\/p>\n<\/div>\n\n\n\n Electrical & Computer Engineering<\/a><\/strong><\/p>\n<\/div>\n<\/div>\n\n\n\n Chair<\/strong>: Dr. Dong Si Due to the rise of mental health issues and shortage of mental health professionals, people turn towards AI powered chatbots and virtual assistants for support and mental health related advice. One such chatbot application is the Data Analysis & Intelligent Systems (DAIS) laboratory\u2019s iCare web application. The Large Language Models used in such chatbots are black boxes and it is difficult to trust and verify their advice. The solution to this lack of transparency is Explainable AI which are tools and algorithms that can provide insight into a machine learning model\u2019s inner workings and explain their decision-making process in a human understandable manner.<\/p>\n\n\n\n This project integrates several Explainable AI algorithms such as Feature Ablations, Layer Integrated Gradients, and Shapley Value Sampling into the iCare web application to explain the LLM\u2019s text generation process. Furthermore, an extensible framework was added to iCare to allow for easy integration of Explainable AI in the future. These algorithms were evaluated based on time and accuracy. User surveys were also conducted to gather feedback on user experience of the explanation feature. Based on the evaluation results, all three algorithms achieved similar levels of accuracy and had excessive processing times. Layer Integrated Gradients performed the best with the highest accuracy and shortest processing time. Additionally, user feedback highlighted a significant preference for natural language explanations over raw token attributions, indicating a need for more intuitive communication of model reasoning.<\/p>\n\n\n\n Chair<\/strong>: Dr. Munehiro Fukuda The demand for distributed data processing grows as modern applications involve increasingly large and complex datasets. Traditional distributed computing frameworks, such as Apache Spark and Hadoop MapReduce, are effective for large-scale data processing but are not always well suited for graph computation. The MASS (Multi-Agent Spatial Simulation) Java library instead provides an agent-based approach to distributed graph computation and has been proven effective for graph computing applications and graph database. However, the performance of MASS Java remains limited in some cases because graph applications often require many agent operations, which introduces significant overhead.<\/p>\n\n\n\n To address these limitations, this thesis introduces several enhancements for improving agent execution performance in MASS Java and evaluates them using graph computing applications and graph database queries. The evaluation shows that the enhancements can improve MASS Java performance in both graph computing and graph database query execution. In addition, this thesis identifies a major overhead in the current MASS graph database and proposes a solution to reduce it. Overall, this thesis contributes to the optimization and evaluation of MASS Java for graph applications and provides useful guidance for future development.<\/p>\n\n\n\n Chair<\/strong>: Dr. William Erdly Pediatric vision impairments frequently go undiagnosed in underserved communities, impacting learning and cognitive development. The EYE Toolbox, developed by the Near Vision Institute (NVI) in partnership with the EYE Research Group at UW Bothell, is a web-based platform supporting NVI’s school-based optometry services across 50+ Washington districts. This project modernized the platform in three engineering phases. Phase 0 hardened 573 PHP files: raw mysqli_query references dropped from 3,905 to 97, jQuery was upgraded from 1.7.1 to 3.7.1, credentials moved to environment variables, and 41 of 45 live black-box attack probes were rejected against the dev deployment. Phase 1 introduced a REST API and React-based frontend to coexist with the legacy PHP\/jQuery pages; paired-endpoint benchmarks showed response payloads 30\u201398% smaller across five surfaces and cumulative session bandwidth 65.1% lower than the legacy path. Phase 2 improved the production RDS vergence therapy application and added three new gamified prototypes (Base Builder, Balloon Pop, Animal Cart) on the shared Phase-1 infrastructure, each preserving the fusion-required vergence demand mechanism. Within-clinic before\/after analysis of NVI’s session telemetry (14,653 sessions, 321 patients) found median peak vergence per session rose from 12.0 to 16.0 prism diopters under the new RDS application (+33% relative), with post-cutover patients leading at 19 of 20 session positions when controlling for therapy-course position; per-session personal-best rate rose from 7.5% to 10.4%. NVI standardized on the new application from cutover forward. A 5-week public demo-portal pilot (199 users, 207 sessions, 16 multiplayer challenges all finishing cleanly) shows voluntary engagement absent clinical referral pressure. A controlled clinical efficacy study integrating all four evaluation pillars is documented for follow-on work.<\/p>\n\n\n\n Chair<\/strong>: Dr. Min Chen The rapid evolution of financial markets demands intelligent trading systems capable of synthesizing heterogeneous information and making adaptive decisions under uncertainty. In this paper, we propose a trading framework that leverages reinforcement learning (RL) to fine-tune a Large Language Model (LLM) for autonomous trade decision-making. Unlike prior approaches that depend on supervised pre-training with expert-annotated analyses or domain-specific corpora for cold-start guidance, our method applies Group Relative Policy Optimization (GRPO) directly to a general-purpose instruction-tuned LLM, enabling the model to develop trading competence purely through reward-driven exploration without curated professional signals.<\/p>\n\n\n\n Themodelingests multi-source market observations \u2014 encompassing technical indicators, financial news, and corporate financial statements \u2014 within a rolling temporal window, and outputs structured trading strategies specifying action type, share quantity, take-profit price, and stop-loss price. This formulation enforces strategy completeness through explicit exit conditions while supporting flexible position sizing, bridging the gap between simplified academic models and practical trade execution.<\/p>\n\n\n\n To guide learning, we design a multi-dimensional reward function grounded in profitability and trading discipline. Each strategy is evaluated on path-dependent profit-and-loss, risk exposure relative to stop-loss levels, position sizing appropriateness, and regulatory adherence, providing fine-grained feedback that cultivates the model\u2019s awareness of both return potential and downside risk.<\/p>\n\n\n\n We conduct comprehensive experiments along three dimensions: (1) model comparisons \u2014contrasting the RL-trained LLM against its base model, alternative LLM architectures, and a DQN-based traditional RL trading agent to quantify improvements from RL fine tuning and LLM-based reasoning respectively; (2) training budget analyses \u2014 investigating how training steps influence the model; and (3) strategy ablations \u2014 examining the contributions of quantity-based position sizing. Results demonstrate that the proposed framework produces coherent, risk-aware trading strategies without supervised warm-up; the ablation analyses further yield insights into the respective roles of model capacity, training sufficiency, and strategy design.<\/p>\n\n\n\n Chair<\/strong>: Dr. Min Chen FinWise is a personal finance web application designed to help users better understand their spending behavior, manage transactions, track financial goals, and receive actionable financial guidance. The system combines a React and TypeScript frontend with a Node.js, Express, and Firebase backend to support transaction management, PDF-assisted transaction import, visual analytics, goal planning, and AI-assisted financial reasoning.<\/p>\n\n\n\n A key focus of the project is explainability. Rather than presenting users with opaque predictions, FinWise combines deterministic financial calculations, machine learning models, and large language model narration to produce responses that are traceable and easy to understand. The goal projection module uses LightGBM-based regression and classification models to forecast monthly savings, estimate goal completion timelines, and evaluate whether users are likely to meet their deadlines. The assistant supports descriptive, predictive, and prescriptive finance questions, helping users interpret trends, compare categories, forecast savings, and explore spending-reduction scenarios.<\/p>\n\n\n\n The project emphasizes accessibility, usability, and financial literacy by presenting complex financial insights in plain language while preserving the underlying calculations. FinWise demonstrates how machine learning and AI-assisted interfaces can be integrated into a practical personal finance tool that supports informed decision-making and goal-oriented financial planning.<\/p>\n\n\n\n Chair<\/strong>: Dr. Min Chen Architectural quality in React applications is difficult to measure and even harder to enforce automatically. As components grow over time, they tend to accumulate responsibilities gradually, through changes that are each individually reasonable, until they become too large and complex to maintain effectively. Unlike syntax errors or type violations, these structural problems are invisible to existing tools such as ESLint, TypeScript, and SonarQube, which means teams rely entirely on code review to catch them, an approach that is inconsistent and does not scale.<\/p>\n\n\n\n This research investigates whether architectural anti-patterns in React can be detected through static analysis with enough precision to be practically useful. The key methodological contribution is a combined metric approach for detecting oversized components, where a component must simultaneously exceed thresholds on Lines of Code, JSX element count, and hook count to be flagged. The reasoning is that any one of these metrics can be elevated for legitimate reasons, but a component that is large, structurally complex, and stateful all at once is almost always doing too much. This achieved 98.04 percent precision in the validation study, compared to roughly 62 percent when using a single metric.<\/p>\n\n\n\n Three independent reviewers assessed 71 components drawn from four production React codebases, Grafana, Mattermost, Refine, and TodoMVC. Average precision across all patterns was 86.39 percent, with Extreme severity cases reaching 100 percent. The 62.5 percent inter-rater agreement also revealed something important: architectural quality is partly subjective, which means the right role for a tool like this is to flag components for human discussion, not to make final judgments.<\/p>\n\n\n\n One unexpected finding was that detection rates were consistent across all four repositories despite them varying enormously in size and age, suggesting that teams naturally maintain a stable level of architectural complexity over time rather than letting debt accumulate indefinitely.<\/p>\n\n\n\n Chair<\/strong>: Dr. Munehiro Fukuda Academic collaboration, citation influence, and institutional research visibility are increasingly reflected through scholarly relationship networks. However, existing academic platforms remain largely profile-centered and do not provide an institution-focused, interactive, and queryable graph system for multi-hop exploration across authors, works, affiliations, and citations. This thesis investigates the design and benchmarking of a UWB citation graph, for practicality, seeded from CSS faculty scholarly activities, and examines how effectively different graph database systems support this richer graph model for practical scholarly exploration.<\/p>\n\n\n\n To address this problem, this work designs and implements a scholarly citation and co-authorship graph pipeline that constructs a heterogeneous Author\u2013Work\u2013citation\u2013Affiliation graph using institutional seed data and OpenAlex-derived metadata. The resulting graph is intended to support practical use cases such as collaborator discovery and referee exploration for UWB CSS faculty. The system is evaluated across three graph databases\u2014Neo4j, ArangoDB, and MASS Graph DB\u2014and is benchmarked using LDBC-aligned workloads and metrics, including bulk ingestion throughput, query throughput, and multi-hop traversal latency. In addition to the institutional citation graph, the evaluation framework includes public benchmark datasets of different graph types, densities, and scales to enable broader cross-platform comparison.<\/p>\n\n\n\n The results show that the heterogeneous scholarly citation graph provides substantially richer analytical capability than simpler single-relation citation graphs by enabling cross-relational exploration over authorship, affiliation, and citation structure. The benchmarking results further indicate that platform strengths are workload-and-graph-dependent: Neo4j performs strongly for interactive read-heavy exploration on the institutional citation graph, ArangoDB remains competitive in selected ingestion-oriented scenarios, and MASS Graph DB performs especially well on loading structurally simpler single-relation graph workloads.<\/p>\n\n\n\n Overall, the findings suggest that richer institutional citation graph modeling can remain practical for interactive exploration, while platform suitability depends on how well each system handles heterogeneous topology, traversal complexity, and deployment conditions.<\/p>\n\n\n\n Chair<\/strong>: Dr. Annuska Zolyomi Social Stories\u2122 are a widely used intervention that helps autistic children understand and prepare for social situations, but creating personalized stories that follow established Social Stories\u2122 criteria can be time-consuming for caregivers. Existing digital tools often provide either fixed story libraries or free-form editors with limited support for methodological fidelity. This project presents Mystoria, an iPad and iPhone application that supports caregivers in authoring personalized Social Stories\u2122 with large language model assistance while preserving caregiver control. Mystoria combines multimodal story creation, including text, AI-generated and camera-sourced images, AAC pictograms, and caregiver-recorded audio, with an AI Draft workflow designed around the structural criteria for Social Stories\u2122. The application includes an in-app draft-quality scorer that helps caregivers review saved drafts against those criteria, as well as a hybrid AI design that combines cloud-based generation with an optional on-device fine-tuned Gemma model for more privacy-aware authoring. Mystoria will be evaluated through a caregiver-only within-subject study in which each participant creates two stories on the same anchor topic: one manually and one with AI Draft. The study examines caregiver preference, trust in AI-generated content, perceived usability and usefulness, analysis of edits and sentence types, and feedback on appropriateness, privacy, and acceptable boundaries for AI assistance. The goal is to understand whether AI-assisted authoring can reduce caregiver burden while preserving personalization, caregiver agency, and adherence to established Social Stories\u2122 criteria.<\/p>\n\n\n\n Chair<\/strong>: Dr. Geethapriya Thamilarasu The rapid expansion of the Internet of Things (IoT) has created the need for anomaly-based intrusion detection systems (IDS) to use Machine Learning. Deep learning models are effective at identifying adaptive security threats, but their opaque nature limits interpretability. Explainable Artificial Intelligence (XAI) addresses this by providing techniques for producing an explanation for a model\u2019s predictions. This project addresses how a hybrid XAI approach can provide accurate, valid explanations for IoT ML IDS models while maintaining real-time performance. The proposed solutions aggregate local explanations for global insight, using efficient feature attribution calculation provided by the FESP method, to produce an explanation using the MAPLE explanation framework that provides local diagnostics with awareness of global patterns. Experimental evaluation demonstrates that hybrid explanation approaches can provide accurate and defensible global and local interpretability while maintaining performance within practical real-time constraints<\/p>\n\n\n\n Chair<\/strong>: Dr. Brent Legesse The maritime logistics industry generates large volumes of operational data from vessel systems; however, much of this data remains underutilized within existing enterprise environments. This work presents the design and evaluation of the MAST (Maritime Analytics and Supporting Technology) platform, a data-driven platform developed in collaboration with Bernert Navigation Inc. to transform raw vessel data into actionable insights that support operational decision-making.<\/p>\n\n\n\n Unlike traditional approaches that require the development of entirely new software ecosystems, MAST leverages existing data infrastructure and augments it with easy to use, scalable capabilities. Through iterative development and stakeholder driven design, the system evolved from an initial focus on performance analytics to a broader platform emphasizing training augmentation and proactive safety improvement. By enabling users to review a large and diverse set of real vessel scenarios, MAST addresses key limitations of conventional training methods, which are often constrained by limited observational opportunities.<\/p>\n\n\n\n Evaluation of the system was conducted through stakeholder engagement, including detailed system walkthroughs and structured feedback sessions with operational leadership. Results indicate strong potential for improving training efficiency and reducing operational risk through increased scenario exposure and data accessibility by supporting a culture of continuous learning. Additional opportunities were identified such as contextual data integration.<\/p>\n\n\n\n This work contributes a practical framework for extending the value of existing maritime data systems while minimizing implementation overhead. More broadly, it demonstrates how data driven platforms can bridge the gap between operational data collection and meaningful organizational impact, enabling scalable improvements in training and long-term performance optimization.<\/p>\n\n\n\n Chair<\/strong>: Dr. Annuska Zolyomi Digital health technologies increasingly help people monitor physical activity, reflect on wellness data, and manage everyday health routines. However, many health-tracking systems emphasize individual metrics such as steps, goals, and progress indicators while overlooking older adults\u2019 accessibility needs, social routines, and preferences for meaningful community engagement. Grounded in Self-Determination Theory, HealthMate investigates how mobile health-tracking technologies can support older adults\u2019 autonomy, competence, and relatedness through accessible design, clear feedback, and socially supportive wellness features.<\/p>\n\n\n\n This work examines older adults\u2019 experiences with physical activity tracking, their preferences for peer-supported and community-based wellness tools, and the tradeoffs they consider when sharing progress or participating in socially oriented health features. Through interviews, co-design sessions, prototype development, and evaluation, the project identifies opportunities for designing health technologies that are usable, motivating, trustworthy, and socially meaningful.<\/p>\n\n\n\n HealthMate supports activity tracking, wellness reflection, community participation, and lightweight social connection. Rather than emphasizing competition or public achievement sharing, the design focuses on practical coordination, accessible progress feedback, and features that help older adults feel confident, independent, and connected. The project contributes design insights for personal health informatics systems that support healthy aging through both individual reflection and shared participation.<\/p>\n\n\n\n Chair<\/strong>: Dr. Munehiro Fukuda MASS (Multi-Agent Spatial Simulation) C++ is a parallel computing library for agent-based simulations on distributed memory clusters, organised around the Bulk Synchronous Parallel model and a master-worker coordination scheme. Three structural limitations have constrained its applicability as a general-purpose runtime. An integer-based method dispatch scheme couples user-defined Place and Agent classes to the framework and propagates renumbering errors silently. A per-iteration master coordination cost of K \u00d7 N barrier round-trips dominates wall-clock time on communication-intensive workloads. And a Places abstraction restricted to regular grids leaves social, biological, and transportation graphs without first-class support.<\/p>\n\n\n\n This thesis presents three feature extensions that address these limitations while preserving backward compatibility with existing MASS C++ programs. A three-tier dispatch architecture replaces integer switch\/case with string-named methods, header-defined lambdas, and JIT-compiled lambdas, unified through a single registry. An IterationConfig phase pipeline backed by an AsyncHandle executor collapses the K \u00d7 N master round-trip pattern into a single dispatch by allowing workers to advance through compute, communication, and agent-management phases autonomously. A graph stack consisting of GraphTopology, GraphPlaces, and GraphAgents extends MASS C++ from grid-only topologies to arbitrary graphs, with bidirectional adjacency, locality-aware partitioning, edge-constrained agent migration, Pregel-inspired combiners and aggregators, and a pluggable parser interface for user-defined graph formats.<\/p>\n\n\n\n The extensions are evaluated on five benchmarks (Wave2D, SugarScape, PageRank, BFS Wavefront, and Random Walk) covering correctness, performance, and programmability. Compound execution reduces barrier round-trips by orders of magnitude and yields speedups that grow with the number of workers on Places-only workloads. SugarScape scaling exposes an O(P\u00b2) bottleneck in the existing all-to-all agent exchange protocol, which caps compound speedup as P grows and motivates the sparse neighbour-rank exchange that GraphAgents adopts. The graph stack also positions MASS C++ as the only system among the surveyed prior art that supports mobile agents over irregular topologies while remaining a drop-in extension of the existing grid-based runtime. The new dispatch tiers reduce per-method boilerplate by approximately 46% at overheads of 0-6% for header lambdas and 5-15% for JIT lambdas.<\/p>\n\n\n\n Chair<\/strong>: Dr. Dong Si Accurate identification of atomic positions and types in macromolecular structures is fundamental to understanding biological function at the molecular scale. Cryo-electron microscopy (cryo-EM) has transformed structural biology by enabling high-resolution visualization of complex biomolecules, yet deriving precise atomic models from cryo-EM density maps remains challenging due to noise, heterogeneity, and variability across datasets.<\/p>\n\n\n\n DeepTracer is an established de novo framework that employs four specialized U-Net models to predict atom locations and types from cryo-EM maps, with each network learning a distinct structural signal that is subsequently integrated into a complete protein model.<\/p>\n\n\n\n With DeepTracer Diffusion, we introduce the next evolution of this framework by integrating diffusion-based generative modeling to improve atomic position prediction and residue classification within cryo-EM density maps. Using a single Denoising Diffusion Probabilistic Model (DDPM), we jointly generate refined atomic coordinates and atom-type labels while remaining fully compatible with the existing DeepTracer post-processing pipeline. To support direct prediction of discrete atom-class labels, we introduce a novel one-hot reverse diffusion procedure that produces segmentation masks at every timestep. Across benchmark datasets, DeepTracer Diffusion increases residue coverage by 24.80\\% and improves the total F1 score by 11.63\\%.<\/p>\n\n\n\n These gains demonstrate that DeepTracer Diffusion more accurately reconstructs atomic structures from experimental cryo-EM maps, even under substantial noise and resolution constraints. Our work advances computational structural biology by providing a scalable, AI-driven framework for macromolecular model building, with broad implications for drug discovery, protein engineering, and structural refinement.<\/p>\n\n\n\n Chair<\/strong>: Dr. Dong Si The iCare platform provides AI-driven mental health support through therapeutic chatbots, yet it currently lacks a systematic method to empirically validate the efficacy and safety of new generative models. To address this bottleneck in digital mental health AI, this project introduces the iCare Validation Engine, a foundational MLOps framework built as an extension of the existing platform. The engine integrates a full-stack A\/B experimentation backend for stable user routing with an automated, asynchronous LLM-as-a-Judge evaluation pipeline that scores conversation transcripts on therapeutic alliance, empathy, and safety. Engineering validation confirmed consistent assignment integrity, correct lifecycle transition behavior, and a score parser success rate exceeding 97% against malformed model outputs. A pilot experiment successfully demonstrated the complete system operating end-to-end, transforming iCare into a dynamic “living laboratory” and establishing a viable foundation for continuous, evidence-based model improvement.<\/p>\n\n\n\n Chair<\/strong>: Dr. Kaylea Champion Prism is a public website that, for each ongoing news event, produces a report showing which specific claims each news source reported, which it omitted, where sources directly contradict one another, and how each article’s tone and framing compare. Existing consumer tools such as Ground News and AllSides answer adjacent questions by attaching bias ratings to whole outlets rather than computing them from each article’s text. Prism’s contribution is to compute those signals from each article’s text at scale, across thousands of sources, without inheriting any third-party editorial labels. Behind the report, Prism extracts atomic claims from every article using a large language model, deduplicates them into a cross-source consensus pool, scores each article against the pool, and flags mutually exclusive claims as conflicts. Tone and framing are computed per article from the text. To establish that the extraction step is trustworthy on unfamiliar single-document text, the pipeline was benchmarked on two academic datasets, reaching F1=0.874 on Rotowire and recall=0.938 on BillSum. New events are ingested and analyzed within minutes of publication. Prism uses “claim” rather than “fact” deliberately: it extracts assertions without evaluating their correctness, and cross-source consensus reflects what is widely reported, not what is true.<\/p>\n\n\n\n Chair<\/strong>: Dr. Munehiro Fukuda Graph representation learning algorithms like Node2Vec generate highly accurate topological embeddings but impose severe memory and computational bottlenecks on centralized architectures. This paper presents a scalable, distributed Node2Vec engine engineered natively within the Multi-Agent Spatial Simulation (MASS) framework. By mapping second-order, biased random walks to autonomous mobile software agents, our architecture efficiently samples complex networks while bypassing single-machine memory limitations. We introduce a novel Compute-Node-Centric training paradigm that pairs isolated Skip-Gram neural network optimization with a decentralized, logarithmic tree-reduction synchronization protocol. Empirical evaluations across benchmark graphs (Cora and OGBL-DDI) demonstrate that this distributed engine achieves strict predictive parity with industry-standard PyTorch baselines across key ranking metrics like MAP, Recall.. etc. Ultimately, this architecture successfully outperforms linear methods like FastRP in topological accuracy while comprehensively unlocking the spatial scalability required for parallelized, deep graph learning.<\/p>\n\n\n\n Chair<\/strong>: Dr. Dong Si Mental health conversational AI systems are becoming more common as tools for emotional support, reflection, and early guidance. For researchers and system designers, transparency helps evaluate whether a system\u2019s responses are grounded in the conversation rather than simply sounding supportive. For users, this transparency matters because they may want to understand whether the system recognized their concern and responded to the right part of what they shared. A common approach is to ask a language model to explain its own response in free-form text, but these explanations can sound reasonable without being closely tied to the model behavior behind the response. This project develops a user-facing explanation feature for CareBot, a conversational AI system within the iCare project at the Data Analysis and Intelligent Systems (DAIS) Laboratory. Instead of generating a free-form justification, the feature traces a selected CareBot response back to earlier user messages that most strongly supported it. The approach uses Layer Integrated Gradients to compute token-level attribution scores for a selected response, maps those scores back to the original conversation, and converts the strongest user-side evidence into short, readable phrases. The explanation feature was evaluated through phrase-level deletion and human evaluation to examine both model-level faithfulness and user perceptions. Results showed that top-ranked phrases had a stronger effect on the model\u2019s response score than low-ranked phrases, while human evaluation helped assess the explanation feature in terms of clarity, usefulness, and transparency.<\/p>\n\n\n\n Chair<\/strong>: Dr. Erika Parsons Traumatic Brain Injury (TBI) is a significant medical condition that can result in long-term neurological impairment or life-threatening intracranial hemorrhage. While computed tomography (CT) and magnetic resonance imaging (MRI) are effective diagnostic tools, their high cost and lack of portability limit accessibility in certain scenarios such as rural and combat environments where rapid triage is essential. Tissue Pulsatility Imaging (TPI), an ultrasound-based technique developed at the University of Washington through Department of Defense funding, offers a potential alternative by enabling the collection of ultrasound data with a portable device. This method measures tissue displacement within the brain resulting from pulsatile blood flow. Previous work at the University of Washington utilized TPI displacement data and CT-derived masks as ground truth to train machine learning models for cranial feature and hemorrhage detection; however, blood segmentation remained challenging due to excessive noise and a limited sample size.<\/p>\n\n\n\n This thesis extends previous research by assessing whether blood-focused segmentation models trained on distinct participant subsets yield different detection outcomes. Eight blood-mode U-Net models were developed using participant groups categorized by scan orientation and blood-region location, with axial and coronal views further divided into top, bottom, left, and right regions. Each model was evaluated on its own held-out test set as well as on the test sets from the other seven groups, enabling comprehensive cross-test comparisons. Participant-level positivity was determined by thresholding predicted blood-mask pixels and applying a majority-vote analysis across models. Support Vector Machine (SVM) and k-Nearest Neighbors (KNN) baselines were also evaluated for comparison with previous results using the entire dataset.<\/p>\n\n\n\n The U-Net models demonstrated limited performance in blood localization, with an average cross-test Dice coefficient of 0.101. Models trained on axial views achieved the highest segmentation accuracy and were the only ones to produce participant-level positive predictions. However, when applying a stricter majority-vote criterion that required at least half of the models to classify a result as positive, no test set met this threshold, as each received only three out of eight positive votes. The SVM and KNN baselines yielded substantially higher positive classification rates, but their results were heavily influenced by class imbalance and lacked spatial blood localization capability. Collectively, these findings indicate that TPI displacement data contain signals relevant to hemorrhage detection, yet reliable intracranial hemorrhage localization remains difficult due to limitations in sample size, diversity, and model architecture.<\/p>\n\n\n\n Chair<\/strong>: Dr. Annuska Zolyomi SafeSpace AI is a human-centered, guardrailed conversational platform designed to provide accessible emotional support for individuals experiencing anxiety, stress, and emotional overwhelm. Existing mental wellness applications often rely on static self-help content, limited personalization, or costly professional support models that may not be readily accessible to students, young professionals, and underserved communities. Furthermore, many AI-powered conversational systems lack emotional safety mechanisms, contextual grounding, and responsible response generation, making them unsuitable for sensitive mental wellness interactions. Unlike conventional wellness applications, SafeSpace AI combines retrieval-augmented generation (RAG), prompt-orchestrated large language models, and multimodal calming interventions to create an empathetic, context-aware, and supportive user experience.<\/p>\n\n\n\n SafeSpace AI employs a retrieval-augmented architecture that grounds conversational responses in curated cognitive behavioral therapy (CBT)-inspired wellness resources, reducing hallucinations and improving contextual relevance during emotionally sensitive interactions. The system integrates guardrailed prompt orchestration, response structuring, and ethical conversational constraints to encourage supportive guidance rather than harmful dependency or unsafe advice generation. In addition, the platform incorporates multimodal wellness tools such as guided breathing exercises, calming sensory interactions, and curated soothing media experiences to provide users with real-time emotional regulation support beyond text-based conversation.<\/p>\n\n\n\n The application further emphasizes accessibility, emotional safety, and human-centered design through a calming interface tailored to reduce cognitive overload and anxiety during use. By combining retrieval-augmented reasoning, conversational AI, multimodal interventions, and ethical guardrails, SafeSpace AI transforms digital mental wellness support from a passive chatbot interaction into an adaptive, supportive, and accessible emotional assistance platform.<\/p>\n\n\n\n Chair<\/strong>: Dr. Geethapriya Thamilarasu The rapid growth of Internet-of-Things (IoT) devices has expanded the attack surface for network-based malware, creating demand for detection systems that are both accurate and interpretable. Machine learning classifiers achieve strong detection performance but provide limited insight into their decisions. Large Language Models (LLMs) offer structured reasoning capabilities, yet applying them directly to numeric network telemetry consistently fails \u2014 and the reasons for this failure have not been systematically studied.<\/p>\n\n\n\n This work introduces a reproducible, geometry-aware benchmark framework that investigates why LLMs succeed or fail at IoT malware detection, and presents the first LLM evaluation on the CIC IoT-DIAD 2024 dataset across both packet-level and flow-level traffic domains. Rather than treating Retrieval-Augmented Generation (RAG) as a performance tool, the framework identifies the class composition of the FAISS (Facebook AI Similarity Search) index \u2014 the pool of historical examples the LLM consults as the primary factor governing classification reliability. A controlled 2\u00d72 experimental design isolates the effects of feature domain and class distribution independently, enabling clean attribution of performance changes to retrieval geometry rather than to model or prompt differences.<\/p>\n\n\n\n The central finding is that retrieval geometry, not prompt design or model capability, governs classification reliability. When real-world traffic is heavily imbalanced, a standard FAISS index becomes geometrically corrupted, causing consistent misclassification \u2014 not because the LLM reasons poorly, but because the examples it retrieves are misleading. This work introduces NatBal (Natural-Balanced Index Conditioning), a correction that rebuilds the FAISS index from balanced training data regardless of test distribution, restoring reliable performance across all evaluation conditions without any model training or fine-tuning. The corrected framework approaches the detection performance of a fully supervised classifier while additionally producing human-readable, neighbor-grounded explanations of each decision.<\/p>\n\n\n\n These results reframe LLM-based intrusion detection as a geometry-sensitive reasoning problem, and provide a fully reproducible reference benchmark for researchers evaluating LLM behavior in network security applications.<\/p>\n\n\n\nMaster of Science in Computer Science & Software Engineering<\/h2>\n\n\n\n
SPRING 2026<\/h3>\n\n\n\n
Monday, May 18<\/h4>\n\n\n\n
Khushaal Kamal Kurswani<\/h5>\n\n\n\n
Candidate<\/strong>: Master of Science in Computer Science & Software Engineering
1:15 P.M.; Join Khushaal Kamal Kurswani’s online defense<\/a>
Project:<\/strong> Building an Extensible Explainable AI Module for Mental Health Conversational AI<\/p>\n\n\n\nTuesday, May 19<\/h4>\n\n\n\n
Bo Fu<\/h5>\n\n\n\n
Candidate<\/strong>: Master of Science in Computer Science & Software Engineering
11:00 A.M.; Join Bo Fu’s online defense<\/a>
Thesis:<\/strong> Enhancing Parallelization of Agent-based Graph Computing<\/p>\n\n\n\nWednesday, May 20<\/h4>\n\n\n\n
Josiah Zacharias<\/h5>\n\n\n\n
Candidate<\/strong>: Master of Science in Computer Science & Software Engineering
3:30 P.M.; Discovery Hall (DISC) 464
Project:<\/strong> Opto-mistic Therapy: Modernizing Stereoscopic Vision Therapy through Cutting-Edge Games<\/p>\n\n\n\nThursday, May 21<\/h4>\n\n\n\n
Dazhi Li<\/h5>\n\n\n\n
Candidate<\/strong>: Master of Science in Computer Science & Software Engineering
8:45 A.M.; Join Dazhi Li’s online defense<\/a>
Thesis:<\/strong> Towards Smarter Trading: An AI Trading Framework Combining Reinforcement Learning and Large Language Model<\/p>\n\n\n\n
\n\n\n\nAqsa Inamdar<\/h5>\n\n\n\n
Candidate<\/strong>: Master of Science in Computer Science & Software Engineering
11:00 A.M.; Join Aqsa Inamdar’s online defense<\/a>
Project:<\/strong> FinWise: Personalized Financial Empowerment<\/p>\n\n\n\n
\n\n\n\nPragnya Ambekar<\/h5>\n\n\n\n
Candidate<\/strong>: Master of Science in Computer Science & Software Engineering
2:00 PM.; Join Pragnya Ambekar’s online defense<\/a>
Thesis:<\/strong> Automated Detection of Architectural Anti-Patterns in React Applications Using Static Analysis<\/p>\n\n\n\nFriday, May 22<\/h4>\n\n\n\n
Yumeng Pang<\/h5>\n\n\n\n
Candidate<\/strong>: Master of Science in Computer Science & Software Engineering
8:45 A.M.; Join Yumeng Pang’s online defense<\/a>
Thesis:<\/strong> Design and Benchmarking of a Citation Graph DB Across Neo4j, ArangoDB, and MASS Graph DB Systems<\/p>\n\n\n\n
\n\n\n\nKris Yu<\/h5>\n\n\n\n
Candidate<\/strong>: Master of Science in Computer Science & Software Engineering
3:30 P.M.; Join Kris Yu’s online defense<\/a>
Project:<\/strong> Mystoria: AI-Assisted Authoring of Personalized Social Stories for Autistic Children<\/p>\n\n\n\nTuesday, May 26<\/h4>\n\n\n\n
Siddharth Thammineni<\/h5>\n\n\n\n
Candidate<\/strong>: Master of Science in Computer Science & Software Engineering
8:45 A.M.; Discovery Hall (DISC) 464
Project:<\/strong> Hybrid XAI-based Intrusion Detection for IoT Networks<\/p>\n\n\n\nThursday, May 28<\/h4>\n\n\n\n
Garrett Woelfl<\/h5>\n\n\n\n
Candidate<\/strong>: Master of Science in Computer Science & Software Engineering
8:45 A.M.; Discovery Hall (DISC) 464
Project:<\/strong> MAST: A Maritime Analytics Platform for Operational Safety and Optimization<\/p>\n\n\n\n
\n\n\n\nHarsha Agarwal<\/h5>\n\n\n\n
Candidate<\/strong>: Master of Science in Computer Science & Software Engineering
11:00 A.M.; Join Harsha Agarwal’s online defense<\/a>
Thesis:<\/strong> Connected Wellness: Designing a Peer-Supported Mobile Platform to Promote Social and Emotional Wellbeing Among Older Adults<\/p>\n\n\n\n
\n\n\n\nRishabh Pratap Singh<\/h5>\n\n\n\n
Candidate<\/strong>: Master of Science in Computer Science & Software Engineering
11:00 A.M.; Discovery Hall (DISC) 464
Thesis:<\/strong> Feature Extension of MASS C++ towards a General Purpose Library<\/p>\n\n\n\n
\n\n\n\nNathaniel Jewel<\/h5>\n\n\n\n
Candidate<\/strong>: Master of Science in Computer Science & Software Engineering
3:30 P.M.; Join Nathaniel Jewel’s online defense<\/a>
Project:<\/strong> DeepTracer Diffusion: Atomic Structure Modeling with Generative Diffusion<\/p>\n\n\n\nFriday, May 29<\/h4>\n\n\n\n
Athresh Kiran<\/h5>\n\n\n\n
Candidate<\/strong>: Master of Science in Computer Science & Software Engineering
8:45 A.M.; Join Athresh Kiran’s online defense<\/a>
Project:<\/strong> The iCare Validation Engine: An Adaptive Framework for Assessing AI Therapeutic Efficacy<\/p>\n\n\n\n
\n\n\n\nAdrian Albu<\/h5>\n\n\n\n
Candidate<\/strong>: Master of Science in Computer Science & Software Engineering
3:30 P.M.; Discovery Hall (DISC) 464
Project:<\/strong> Prism: Article-Level Media Comparison Through Automated Claim Extraction<\/p>\n\n\n\n
\n\n\n\nDeepak Sujay Gudiseva<\/h5>\n\n\n\n
Candidate<\/strong>: Master of Science in Computer Science & Software Engineering
5:45 P.M.; Discovery Hall (DISC) 464
Project:<\/strong> Agent-Based Distributed Node2vec<\/p>\n\n\n\nMonday, June 1<\/h4>\n\n\n\n
Meghana Dayathri<\/h5>\n\n\n\n
Candidate<\/strong>: Master of Science in Computer Science & Software Engineering
8:45 A.M.; Join Meghana Dayathri’s online defense<\/a>
Project:<\/strong> Model-Grounded Explanations for Responses Generated from Conversational AI<\/p>\n\n\n\nTuesday, June 2<\/h4>\n\n\n\n
Enbai Kuang<\/h5>\n\n\n\n
Candidate<\/strong>: Master of Science in Computer Science & Software Engineering
1:15 P.M.; Join Enbai Kuang’s online defense<\/a>
Thesis:<\/strong> Application of Deep Learning for the Detection of Intracranial Hemorrhage Through Different Planes Using Ultrasound Imaging<\/p>\n\n\n\nThursday, June 4<\/h4>\n\n\n\n
Sonal Singh<\/h5>\n\n\n\n
Candidate<\/strong>: Master of Science in Computer Science & Software Engineering
11:00 A.M.; Join Sonal Singh’s online defense<\/a>
Project:<\/strong> SafeSpace AI: A Human-Centered Guardrailed Retrieval-Augmented Conversational System for Anxiety and Stress Support<\/p>\n\n\n\nMaster of Science in Cybersecurity Engineering<\/h2>\n\n\n\n
SPRING 2026<\/h3>\n\n\n\n
Tuesday, May 26<\/h4>\n\n\n\n
Suryanarayana Putrrevu<\/h5>\n\n\n\n
Candidate<\/strong>: Master of Science in Cybersecurity Engineering
11:00 A.M.; Commons Hall (UW2) 327
Project:<\/strong> LLM-Based IoT Malware Detection via Similarity-Grounded Context<\/p>\n\n\n\nWednesday, May 27<\/h4>\n\n\n\n
Harsh Makarand Jannawar<\/h5>\n\n\n\n