Capstone Projects

Abstract
A majority of enterprise dialog systems today
are built upon finite-state dialogue manage-
ment architectures, which enforce rigid, rule-
based call flows. These systems often strug-
gle to handle evolving user goals or recover
gracefully from unexpected conversational in-
puts. In contrast, human agents excel at main-
taining context, reasoning about what infor-
mation is needed, and choosing whether to
ask questions or consult external tools. This
project explores the design of agentic large
language models (LLMs) that emulate these
human capabilities. We propose a dialog agent
that integrates multi-turn reasoning, tool use,
and adaptive decision-making to dynamically
manage conversations. Inspired by frame-
works like ReAct, ReSpAct, and Pre-Act, our
agent determines whether to prompt users or
take autonomous actions based on situational
context. To support development, we gener-
ate a synthetic dataset using dual-LLM simu-
lations, representing realistic, annotated dialog
traces with intermediate reasoning steps. Our
system is designed to generalize across task-
oriented domains, supporting both conversa-
tional and non-conversational workflows. This
work contributes toward more flexible, user-
aware AI agents that improve automation qual-
ity and reduce friction in real-world human-
computer interactions.

Abstract
As the volume of academic publications con-
tinues to grow exponentially, researchers face
increasing difficulty in both discovering rele-
vant literature and evaluating its reliability. Our
proposed project, No RAGrets, will introduce a
modular NLP system designed to assist with lit-
erature discovery, information extraction, and
early-stage paper vetting. The system will be
aimed at helping users identify key research
claims, assess citation quality, and surface po-
tential red flags such as methodological incon-
sistencies or weak references.
We will approach this challenge by combining
components for information extraction, seman-
tic classification, and contextual retrieval. Addi-
tionally, we explore the integration of retrieval-
augmented generation (RAG), large language
models (LLMs), and agent-based architectures
to support flexible, interpretable research work-
flows. Our goal is to build a system that not
only improves the precision of literature search,
but also enables more critical and structured
engagement with scientific texts.

Abstract
Our primary goal is to develop an LLM-
based application that supports students during
the writing process. Drawing on existing tools
like Grammarly and WordTune as inspiration,
we envision a system that offers real-time sug-
gestions or conducts a final review after writing
is complete. However, unlike these tools, our
application is designed to preserve cognitive
engagement by prompting rather than correct-
ing. For instance, when grammatical or fluency
issues are detected, the LLM would avoid imme-
diately supplying the fix. Instead, it would ini-
tiate a guided sequence: first asking the student
whether they believe any changes are needed,
then prompting them to identify and explain
those changes. If the student is on the right
track, the LLM could assist in completing the
revision. If not, the system would scaffold the
process further—helping the student recognize
the error, understand the reasoning behind the
correction, and apply the fix themselves. This
approach aims to maintain the learning bene-
fits of writing while still leveraging the LLM’s
strengths.

Abstract
International enterprises typically store data in
a relational database that is accessible via a
dialect of SQL. The barrier-to-entry for deriv-
ing insights from these databases is a need for
both English and SQL fluency. As a means
to simplify data access, we seek to research
text-to-SQL that converts multilingual natural
language questions into SQL queries. While
there have been efforts to increase the presence
of multilingual datasets, traditional methods of
human-translation can be costly and imprac-
tical. Existing synthetic data generation tech-
niques in this realm are limited and restricted
only to English, with filtering processes that
may reduce the difficulty of the SQL queries.
There have also been successful strides towards
grounding text-to-SQL LLMs through the injec-
tion of contextual information, however, these
models still struggle with providing sensible,
complex, and diverse outputs. In addition, ex-
isting datasets frequently overlook cultural and
practical usage norms. For instance, they may
assume that column names follow English con-
ventions without justification, or that users in-
teract with applications using native scripts,
even in contexts where Romanized versions
are more common. This mismatch can lead
to unrealistic training data. To address these
challenges, we propose a method for multilin-
gual synthetic data creation and use it to cre-
ate a novel multilingual dataset for NL (natu-
ral language) to SQL tasks, expanded to Ara-
bic, Chinese, French, German, Hindi, Japanese,
Malay, Portuguese, Russian, Spanish, Turkish,
and Vietnamese. Prior research frequently com-
bined machine translation with human evalua-
tion. With the increasing fluency of LLMs, we
see the opportunity to fully automate the trans-
lation process in cases where we are unable to
utilize human-constructed translations. To our
knowledge, this would be the largest multilin-
gual text-to-SQL dataset and the first scalable
synthetic data generation pipeline for this task.

Abstract
The Samsung GUI Agent project aims to deliver (1) a browser-extension capture engine to record
high-fidelity user actions (clicks, scrolls, inputs, DOM snapshots) and state context (screenshots, ac-
cessibility trees); (2) a high quality dataset including synthetic (MiniWoB++, ALFWorld), realistic
(WebArena, REAL, WebGames), and enterprise (WorkArena, WorkArena++) benchmarks; and
(3) LLM-based agents fine-tuned and evaluated in BrowserGym to exceed state-of-the-art perfor-
mance across these benchmarks. Success metrics include end-task success rate, time to completion,
accuracy of individual action decisions, and robustness under dynamic web changes.

Abstract
As large language models (LLMs) and vision-
language models (VLMs) continue to advance,
there is growing interest in developing GUI
agents capable of interacting with graphical
user interfaces to automate complex tasks
across consumer and enterprise applications.
While significant progress has been made in
consumer-facing applications, enterprise de-
ployment of GUI agents presents unique chal-
lenges including legacy system integration,
stringent security requirements, compliance
considerations, and the need for robust er-
ror handling in mission-critical environments.
This proposal highlights the evolution of LLM
agents, provides a short survey of related work,
and proposes research directions for our cap-
stone project. Our work aims to bridge the gap
between academic research and practical en-
terprise deployment by developing an agentic
system specifically designed for Adobe’s enter-
prise GUI automation challenges, focused on
the Adobe Express application.