Abstract
Technical Project Abstract:
G.L.A.S.S. (Glasses for Linked Automation of Smart Systems) is a wearable smart-glasses control platform designed to let users interact with multiple smart devices through vision-based device recognition and hand-gesture input. The project was motivated by the need for a more intuitive, unified, and accessible way to control household electronics and other connected technologies. Rather than requiring separate remotes or interfaces for each device, G.L.A.S.S. provides a single wearable interface that allows a user to look at a device, confirm the selection, and issue commands through gestures.
The system architecture combines two Raspberry Pi Zero 2 W modules mounted on the glasses with a central Raspberry Pi 5 controller. One onboard unit handles camera streaming and button input, while the other drives a micro-OLED display that gives the user feedback about detected and connected devices. The central controller performs device recognition, gesture interpretation, state management, and communication with target devices over Wi-Fi or Bluetooth. To support recognition of individual devices without retraining a full classifier for every new object, the project uses image embeddings and a vector database to identify devices by nearest-neighbor similarity. This design allows new devices to be added more flexibly than with a conventional closed-set classifier.
The prototype was tested on a smart light, a smart TV, and a computer mouse. Results showed successful end-to-end control across all three devices, object-detection accuracy of 97.7% on a validation dataset, and an average detection time of approximately 0.21 seconds per image. User feedback indicated that the system was comfortable, usable, and effective, though somewhat heavy. Overall, G.L.A.S.S. demonstrates the feasibility of a wearable, gesture-based universal interface for smart-device control, with promising implications for accessibility, convenience, and future expansion.
STS Paper Abstract:
This paper examines whether AI agents are truly becoming democratizing technologies that lower barriers to entry into labor and free people to focus on more meaningful work. Public discourse often presents AI agents as autonomous systems capable of replacing workers, transforming industries, or allowing non-experts to perform highly skilled labor. To investigate these claims, this project analyzes public documentation, product messaging, and user accounts of several AI-agent and AI-assistant technologies, with particular attention to coding agents such as Bolt.new and note-taking assistants such as Fireflies, Otter, and similar platforms. Using Actor-Network Theory (ANT) as an analytical framework, the paper studies how these systems do or do not stabilize as actors within labor networks.
The central finding is that AI agents rarely succeed as durable, autonomous participants in labor networks. Systems marketed as independent workers tend to fail when they are treated as central coordinators of work. In coding contexts, Bolt.new was frequently associated with abandoned projects, debug loops, ignored instructions, and heavy dependence on human intervention, especially among non-technical users. In contrast, more successful AI systems were assistants rather than agents: they performed narrow, low-risk tasks while remaining under human supervision. Coding assistants showed some productivity benefits for expert users, while note-taking assistants persisted only when users treated them as optional and supplementary rather than fully reliable.
This paper argues that current AI agents are not democratizing labor in the broad sense often promised by public rhetoric. They do not substantially remove barriers for average users, and they are not yet displacing humans in stable labor roles. Instead, limited benefits tend to accrue to already-skilled workers who can supervise, correct, and strategically deploy AI tools. AI systems therefore enter labor networks most successfully when they remain peripheral, weakly enrolled actors rather than indispensable ones.
Connection Between Technical and STS Projects:
My technical and STS projects both explore the real relationship between humans and AI systems rather than the idealized one often presented in public discourse. In my capstone project, G.L.A.S.S., the goal was to build a practical wearable system that uses computer vision and gesture recognition to help users control smart devices more intuitively. That project required thinking carefully about reliability, usability, and how much autonomy an AI-enabled system should actually have before users lose trust in it. My STS paper examines a similar issue from a societal perspective by asking whether AI agents can function as meaningful actors in labor networks. Together, the two projects show that successful AI systems are rarely fully autonomous replacements for humans. Instead, they are most effective when designed around human oversight, clear feedback, and constrained tasks. Working on both projects showed me that AI systems are useful only when carefully integrated into human workflows, rather than treated as replacements for human judgment.
