The Sprout is a Unity-based, Steam-ready RPG developed as a one-year capstone project. It demonstrates that a chapter-based RPG demo can be designed, built, publicly released, tested by real players and iteratively improved under practical release constraints. 

The game’s narrative concept was developed by me, and all in-game art assets were hand-drawn by me to maintain a consistent visual identity. From a software engineering perspective, the game is supported by a modular architecture featuring persistent managers, EventBus communication, MVC organization and a data-driven localization pipeline, together with Steamworks SDK integration for Steam deployment. 

The demo is currently available on Steam and supports both English and Chinese. 

Large language model (LLM) inference is increasingly constrained by latency, making low-overhead execution an important systems objective. Existing serving systems typically follow a kernel-per-operator design, where computation is decomposed into many GPU kernels separated by coarse synchronization points. This execution model incurs repeated kernel launch overhead, limits fine-grained control over scheduling and synchronization and prevents effective overlap across operators.  

We propose Mirage Persistent Kernel (MPK), which addresses these limitations by compiling model inference into a single persistent mega-kernel. Instead of treating operators as the unit of execution, MPK represents computation and communication as fine-grained SM-level tasks connected by explicit dependencies and executes them using an in-kernel event-driven runtime. This design eliminates repeated kernel launches, enables fine-grained scheduling and synchronization within a single kernel and exposes opportunities for cross-operator software pipelining and compute-communication overlap that are difficult to realize in conventional systems. The broader MPK project shows that this approach can substantially improve LLM serving performance while preserving a usable compiler-driven workflow.  

Beyond the MPK, this thesis also presents how to extend MPK to multi-GPU low-latency inference. To conform to the persistent megakernel paradigm, I design and implement a set of one-sided, GPU-initiated collective communication mechanisms that integrate directly with MPK’s task abstraction and event-driven runtime. This design preserves a uniform execution model across single- and multi-GPU execution, while enabling communication to compose naturally with fine-grained scheduling and overlap inside the megakernel. The thesis further analyzes the design insights imposed by this setting, including the tradeoffs among communication granularity, synchronization and compute-communication overlap. 

Blockchain technologies are increasingly implemented and used in real-world systems, but they still face many of the same vulnerabilities found in traditional software. Public exploits have shown that insecure smart contract development may cause significant financial risks. Documented losses have reached billions of dollars in recent years. Despite this, there is a lack of hands-on educational resources for learning smart contract security. Graduate-level students (soon-to-be developers) commonly rely on automated analysis tools. These tools are not designed for learning and are known to miss common vulnerabilities.  

This research intends to teach secure smart contract development by developing a curriculum that instructs students to intentionally secure vulnerable smart contracts through Capture-the-Flag (CTF) challenges. The curriculum is stored in an open-source GitHub repository containing vulnerable Solidity contracts. Each of these contracts is paired with learning objectives, background knowledge for the base lesson, pre- and post-quizzes and mitigation strategies. This allows learners to progress from simply using an automated tool to understanding possible attack vectors and how to fix those vulnerabilities before they occur. Common vulnerabilities such as access control, integer overflow, front running and reentrancy are emphasized to teach the foundations of common security flaws in smart contracts.  

The curriculum integrates CTF challenges that promote active problem-solving and Experimentation in a controlled environment. These problems leverage the principles of active learning and experiential education. In addition to developing smart contract technical skills, the curriculum encourages students to think critically. It helps them apply secure programming practices found in traditional software to their smart contract development.  

This curriculum is tested through a pilot study with 3–10 graduate-level students to assess its effectiveness, learners’ comprehension of the material and their confidence in developing secure smart contracts without an automated tool. The results of this study will inform further refinement and the adoption of the curriculum in a larger study across graduate-level courses

Malware analysis through reverse engineering remains a labor-intensive process. Analysts often need to gather and connect evidence from many different tasks, including static binary analysis with tools such as Strings, FLOSS and Binwalk; examination of disassembly and decompiled code in platforms like Ghidra, IDA and Binary Ninja; and dynamic analysis using virtual machines, debuggers and system or network monitoring tools. While large language models (LLMs) have become increasingly capable of understanding code, generating text and invoking tools, it is still unclear how effectively they can assist with binary analysis, where evidence is distributed across specialized workflows rather than contained in source code alone.  

This challenge is especially important in malware analysis, where samples often use evasion and obfuscation techniques designed to hinder investigation. Malware may conceal key strings or imports, detect observation environments, alter behavior at runtime or dynamically generate code in order to obscure its true functionality. This thesis examines whether tool-orchestrated agentic workflows can improve reverse engineering performance, with a particular focus on malware analysis. Rather than relying on a single monolithic agent, the system uses a multi-agent architecture composed of planners, specialized workers, validators and reporters. These agents are given structured access to tools through Model Context Protocol-style integrations, including Ghidra, static-analysis utilities, unpacking support and other artifact-recovery workflows. The research studies how staged reasoning, explicit tool use and evidence review affect tasks such as behavior reconstruction, configuration decoding, packing detection and malware triage.  

To support both practical use and evaluation, this thesis presents two core interfaces built on top of the underlying workflow. The first is a web-based chat interface that provides live pipeline visualizations, tool logs, planner task output, progress tracking and advanced settings for decisions such as pipeline configuration, concurrent worker execution and validator strictness. The second is a testing harness for systematic evaluation, supporting paired tasks and samples, baseline and experimental sweeps, repeated runs, structured judging and graph generation for Analysis.  

The contributions of this work are both systems-oriented and empirical. On the systems side, it offers a modular architecture for agentic reverse engineering that treats tool use and evidence-backed claims as first-class concerns. On the empirical side, it investigates which workflow choices meaningfully improve performance, which mainly affect style or cost and where orchestration itself becomes a limiting factor. More broadly, this thesis aims to clarify what would be required for LLM-based systems to become more reliable assistants for reverse engineering.  

Large language models (LLMs) have attracted broad attention since their emergence and have been widely integrated into daily workflows. However, hallucination, the phenomenon that LLMs generate responses that appear plausible but are factually incorrect or misaligned with user intents, remains a critical challenge to their utility and reliability. Existing hallucination detection methods typically focus on detecting a single dimension of hallucination, such as factuality or faithfulness, which limits their ability to yield comprehensive assessment with clear explanations.  

We propose MACHA, an efficient self-adaptive multi-agent framework for comprehensive hallucination assessment through structured multi-agent debate and multi-dimensional evaluation. MACHA performs parallel evaluation of both factuality and faithfulness and produces reliable hallucination predictions with reasoning for improved interpretability. In addition, an online agentic memory module is incorporated to maintain a trajectory store that serves as the source of self-generated in-context learning examples for future evaluations. The system also periodically reflects on recent success and failure patterns to enable automatic context adaptation, allowing it to improve continuously without additional training.  

Across six benchmarks covering different hallucination detection settings, MACHA improves the average success rate by 5–10% over the base model, while the online context adaptation mechanism further demonstrates promising generalization ability. 

Cross-modal self-supervised learning has shown strong potential for learning representations for audio-visual emotion recognition (AVER). However, most existing methods operate entirely in continuous embedding spaces, which can inadvertently preserve irrelevant patterns, spurious correlations and modality-specific noise. To filter noise while preserving task-relevant semantic information, we propose the Hierarchical Quantized Cross-modal Masked Autoencoders (HQC-MAE) framework.  

This method integrates discrete representation learning into masked data modeling and contrastive learning via learned codebooks, creating hierarchical information bottlenecks in a multimodal setting. We perform cross-modal contrastive learning in a discrete space, encouraging semantically aligned audio and video samples to match in a structured embedding space. Our vector quantization modules are integrated into pretrained unimodal foundation models (VideoMAEv2 for video and AudioMAE++ for audio), enabling us to leverage large-scale pretrained unimodal models rather than training them from scratch. Experiments on emotion recognition benchmarks show consistent improvements over AVER baselines using continuous representations.  

While Retrieval-Augmented Generation (RAG) pipelines address the limitations of large language models (LLMs) by utilizing external knowledge bases, they also introduce new attack surfaces at the retrieval layer. This thesis implements defense mechanisms that leverage Intel SGX-based Trusted Execution Environments (TEE), a hardware-based isolation mechanism and evaluates their security-utility trade-offs and performance overhead. Unlike previous studies that focus on individual defense mechanisms, this thesis comparatively evaluates multiple defense mechanisms within a unified experimental environment and quantifies their effectiveness against each attack category.  

The framework features five modular defense tools placed in the SGX enclave: attested update verification (ensuring updates come from trusted sources), provenance filtering (screening sources of retrieved data), access control (regulating user permissions), document-level encryption (protecting documents individually) and output filtering (checking outputs for security breaches), each independently configurable. This thesis evaluates four categories of attacks under two adversary levels (naive and adaptive): data poisoning, prompt injection, membership inference (determining whether specific documents exist in the retrieval database) and query pattern inference. These attacks are evaluated on an 11.5 million-document dataset in seven configurations, including a baseline and a bundled setting.  

The framework is implemented and evaluated on production Intel SGX hardware. Experiments show that TEE-based defenses substantially mitigate naive attacks, particularly those targeting data provenance, while providing more limited protection against adaptive adversaries that exploit model-level vulnerabilities. This thesis also finds that system prompt design significantly affects resilience against prompt injection, suggesting that prompt hardening and TEE-based defense mechanisms serve as complementary defense layers. Finally, the framework's limitations are discussed, particularly concerning supply-chain poisoning and metadata-based inference, along with directions for strengthening these gaps in future work. 

Spoken Language Identification (LID), the task of automatically determining which language is spoken in an audio recording, is a foundational component of multilingual speech processing systems. Despite advances in self-supervised learning and multilingual speech models, robust LID remains challenging due to domain mismatch, limited low-resource data, dialectal variation and code-switching. 

This thesis addresses these challenges through three contributions. First, we investigate strategies for adapting multilingual speech foundation models under limited and imbalanced data conditions, achieving second place in the Interspeech 2025 ML-SUPERB 2.0 Challenge. Second, we propose geolocation-aware LID, a framework that incorporates geographic information to improve robustness to dialectal and accented variations, achieving state-of-the-art accuracy on FLEURS and ML-SUPERB 2.0. Third, we introduce CS-YODAS, a 313-hour corpus of naturally occurring code-switched speech, enabling zero-to-one improvement in code-switched LID. Together, these contributions advance LID robustness across languages, dialects, accents and real-world multilingual settings.