I am a solution architect currently pursuing a part-time PhD in Information Science at Drexel University. Initially interested in the combination of computer science and linguistics during undergrad, I am particularly interested in how NLP can simplify data analysis/comprehension. I have a secondary interest in how machines can better comprehend human language idiosyncrasies and how we can leverage data to address social issues.

On a personal note, I enjoy hiking, running, and playing open world video games.

Temporal Segmentation of Social Media Timelines: A Comparative Evaluation of Clustering Algorithms

2025 12th International Conference on Social Networks Analysis, Management and Security (SNAMS)

This work evaluates the effectiveness of clustering algorithms for temporally segmenting social media timelines into distinct episodes of use. Using only post timestamps, MeanShift, K-Means, and DBSCAN clustering methods are compared against human annotations that leverage additional metadata such as post content, device type, and language. Performance is assessed through a multi-faceted evaluation including: (1) alignment with human-labeled segmentation points to measure accuracy; (2) runtime comparisons to assess algorithmic efficiency; and (3) analysis of intra- and inter-cluster token similarity to validate the semantic coherence and distinctiveness of the resulting segments. Results show that Mean-Shift clustering consistently produces segments that closely match human annotations, while maintaining computational efficiency and yielding high internal lexical consistency. These findings suggest that timestamp-based clustering provides a low-dimensional, reproducible method for detecting episodes of social media use. This segmentation supports downstream tasks such as author attribution and verification by providing a foundation for longitudinal analyses of online behavior and construction of sequential datasets within a single user account

Authorship in Motion: Stylometric Drift Aware Authorship Attribution in Social Media

In Progress

Author attribution models achieve strong empirical performance in traditional text classification, yet they are largely brittle when applied to the dynamic, continuous stream of social media; existing systems treat account history as a single, static corpus, failing to capture normal stylistic drift or sudden shifts in account ownership. This limitation is particularly problematic in domains such as digital forensics and bot detection, where treating diverse digital identities as aggregate classes obscures crucial feature nuance and leaves platforms vulnerable to deceptive account manipulation. This dissertation investigates how intra-account stylochronometry can be modeled through temporal segmentation to detect potential changes in account operators—or regime changes—without sacrificing classification accuracy. The work progresses from temporal data preprocessing to comparative feature evaluation, and ultimately to a multi-model benchmark for binary anomaly detection. RQ1 investigates how temporal segmentation and behavioral clustering can be optimized to model an individual account's baseline stylistic drift while preserving critical feature nuance. RQ2 evaluates the extent to which specific lexical, syntactic, and structural stylometric features vary in stability when an account shifts from a human operator to an automated system or a different human entity. RQ3 examines how large language model (LLM) approaches compare to traditional machine learning classifiers in terms of accuracy and computational efficiency when performing a binary determination of an account regime change. The expected contributions to the fields of data science and computational linguistics are threefold: (1) a novel temporal framework for intra-account authorship analysis that captures natural stylistic drift over time; (2) an optimized feature selection and preprocessing methodology that establishes benchmarks for stylistic stability in short-form, high-velocity social media environments; and (3) an empirical evaluation of traditional and LLM-based architectures for detecting account transitions which may be due to account compromises, bot activations, and ownership transfers.

Principal Solution Architect / Comcast

September 2021 - Current

Entertainment Metadata Architect supporting apps on Sky, NBC, and Comcast platforms. Focused on Sports, Music, and VOD metadata. Supporting inbound and outbound metadata flows, architecture work revolves around a centralized master data management (MDM) platform for entertainment metadata. The metadata empowers complex content discovery and personalization use cases.

Solution Architect / Cigna

July 2013 - September 2021

Career Progression from: Java Developer, Technical Lead, Delivery Lead, Senior Systems Analyst, to Solution Architect in the Provider Data Domain. Designed large scale solutions involving multiple custom and vendor applications providing users the tools to analyze and manage provider data. Working in a quasi-researcher function, evaluated vendors, produced options, and worked with the scrum teams to deliver the final solutions.

604.744 Information Retrieval Grader / Johns Hopkins University

August 2018 - Dec 2019

Supported a course focused on efficient storage, organization, and retrieval of information

Investigating Diversity in Synthetic Datasets Generated by LLMs

Synthetic datasets are increasingly used to reduce the cost of human data collection and labeling, and to address gaps in dataset availability. While prior work has shown that models trained on synthetic data can underperform compared to those trained on human-generated data, the reasons for this gap are debated. Li et al. attribute the issue to the lack of nuanced, representative samples; other researchers point to a lack of diversity in synthetic data. This project focuses on empirically studying the diversity hypothesis.

Episodes of Social Media Use and Evidence of Regime Changes

Utilizing a repository of 8+ year old Twitter accounts, this research includes developing a taxonomy for episodes of social media use, algorithmically segmenting timelines into episodes of use, evaluating the algorithmic segmentation. Finally features will be identified to train a model on predicting if the author of posts within an episode of social media use is the same author of a previous episode of use - this makes it a low context author attribution task.

Automated Sarcasm Detection within Chat Bot conversations

Customer service interactions are currently being steered towards chat bot interactions, but chat bots are not skilled at detecting the use of sarcasm or irony in responses. Looking at current methods of sarcasm detection on stand-alone tweets and customer reviews, then applying methodologies conversations (expanding datasets to reply-tos and customer service logs).

Bad Data Propgation through Data Networks

Theory: Valid data updates within an integrated domain should behave differently, possibly entering the network through multiple nodes, than bad data updates. i.e. Fake events or facts should propogate through news sites in a different manner than a real event

Poet Author Attribution

When identifying the poet for a given stanza of text, are morphological/phonological/style statistics more relevant than word choice?

Hiking Ontology

Hiking Trail Selection can be a convoluted process for hikers unfamiliar with available trails. Individuals may seek to find trails of specific length, location, or other feature, but while this information is available in the descriptions of trails on local websites, the information is neither linked nor searchable. The Hiking Trail Ontology seeks to address this gap.

NFLTwitterBack

A database of NFL Running Back game performance metrics and social media activity metrics using the SportsDataIO API and Twitter API. Preliminary structure to support questions around the impact of social media use/trends on player performance.