48440 research outputs found
Sort by
Design and Test of a Transcutaneous Oxygen Sensor Prototype
No full textNeonatal monitoring is a significant issue for observing the health of preterm infants. Current methods, such as the Pulse Oximeter, are not reliable for measuring changes in partial pressure of oxygen, which is vital to determine breathing fluctuations in the infant. The Integrated Circuits and Systems Lab (ICAS) at WPI has developed a prototype device that can accurately measure transcutaneous partial pressure of oxygen, but is suitable for adult monitoring for its large size. The purpose of this study is to take the Transcutaneous Oxygen Monitor (TOM) Rigid design and to compact/modify the circuit board to 27 mm in diameter. The board will be split into two separate circuit boards, each performing different functions that communicate through a ribbon cable. The end goal is to prove that a smaller design is feasible
CS MQP - Summarization via LLM CloudBees
No full textThis project addresses the challenge of maintaining unit tests in rapidly evolving code bases by developing a large language model (LLM) tool that analyzes new code changes using GitHub Actions to identify areas needing additional testing. It generates predictive test cases and creates pull requests with ready-to-use code, enhancing developer efficiency. Ultimately, this solution aims to improve software quality by enabling developers to focus on code rather than test development. We used Langflow, Python, YAML, and GitHub Actions to develop a multi-agent workflow. It embeds and stores the entire repository in a vector database using Retrieval-Augmented Generation (RAG). The agents generate test cases, download required dependencies, and create additional test cases iteratively to reach a certain coverage threshold that the developer can set. Our tool achieved code coverage as high as 95% on Cookiecutter, a medium-sized open-source Python repository
Return on Investment: Converting Efficiency into Profitability
No full textWithin the modern world, optimizing operational efficiency and productivity is crucial for any process. Tool organization companies play a key role in making these goals achievable. Sonic Tools USA, a premier tool provider for the automotive, aviation, and manufacturing industries, is seeking to bridge the knowledge gap of potential customer investment certainty. This project empowers Sonic’s clients with data-driven insights, enabling informed decisions and identifying opportunities for long-term improvements
Finite entropy solutions for scalar conservation laws
No full textFor this project, we investigate finite entropy solutions to the one-dimensional Burgers’ equation. We focus on understanding a question connected to how finite entropy solutions deviate from classical entropy solutions. We answer the question in a special case. We also try to get some insight into general properties of finite entropy solutions
Roof Sheathing Press Prototype: Phase 2
No full textThis project aims to address plastic waste and high roofing costs in rural Ghana by developing a coir-reinforced polyethylene terephthalate (PET) composite roofing material using locally available resources and methods. Coconut fiber (Coir) and PET, a common plastic waste product, were chosen for their abundance in the region. Through co-design, small-scale compression molding, and local heating solutions, prototype tiles were produced. Findings indicate that heat control and fiber layering prevent cracks when subject to force, improving tile durability. We recommend further research into optimizing heat sources, scaling up production through improved molding techniques, and conducting a cost-benefit analysis to assess economic feasibility
GraphRAG vs. RAG: Comparative Evaluation of LLM Performance
No full textThe increasing complexity of data in alternative investing demands innovative solutions. TPG Angelo Gordon currently employs Retrieval-Augmented Generation (RAG) for information retrieval and document ingestion in their AI chatbot. While effective, RAG struggles to answer complex queries requiring synthesis across documents. To address this, we evaluated Microsoft’s GraphRAG, which utilizes knowledge graphs to achieve more sophisticated reasoning and contextually relevant responses. We extensively tested and evaluated RAG and GraphRAG across multiple datasets provided by TPG Angelo Gordon using DeepEval, an open-source LLM evaluation framework. Results were synthesized using Snowflake and SnowSQL and visualized in Power BI
Runtime Efficient for Deep Neural Networks
No full textWith deep neural networks (DNNs) being increasingly deployed on low-resource devices, achieving efficiency without sacrificing accuracy has become a critical challenge. Model pruning, commonly performed during training, is a common approach to reducing model size and computational cost, even in environments where computational resources are not limited. However, traditional pruning methods are inherently static, often leading to accuracy losses that hinder their applicability in dynamic, real-world settings. Although some inference-time pruning techniques exist, these methods are generally structured, introducing similar accuracy compromises as conventional pruning methods. In this paper, we propose a novel, unstructured pruning algorithm that performs adaptive, input-specific pruning during inference: Unstructured Inference-Time Pruner (UnIT Pruner). Unlike traditional approaches, UnIT Pruner dynamically skips redundant operations in real time based on the unique properties of each input, allowing for efficient computation without significant accuracy trade-offs. Our algorithm can complement existing methods, enhancing the balance between energy efficiency and accuracy and, in certain systems, even reducing latency. Experimental results show that our method reduces MAC operations by 70.38--87.39% with as low as 0.43% accuracy drop. Additionally, we achieve 74.4--94.7% less inference time and consume 74.2--96.5% less energy on microcontrollers when compared to other pruning methods. Our proposed algorithm achieves state-of-the-art inference-time pruning performance, demonstrating its effectiveness in deploying DNNs in resource-constrained environments
Developing Educational Labs for Reverse Engineering IoT Devices
No full textEducation addressing IoT reverse engineering is currently lacking at WPI and other universities. This paper details the creation of curriculum designed to address this gap in education. Using tools common to the field of IoT reverse engineering four labs were created that each address an important stage of the reverse engineering process. A longer project to be done in parallel with the labs was also developed. Additionally, this paper details ways to improve upon its work in the future
AR Glasses: Application and Design for Education
No full textThis study examines how heads-up augmented-reality (AR) widgets can address persistent challenges in live lecture delivery, namely pacing, personalized student-professor engagement, and increasing cognitive load for recall, by overlaying key instructional aids directly in the instructor’s field of view. Building on our semi-structured interviews and classroom observations with computer-science faculty, we applied axial coding and affinity mapping to pinpoint three core pain points: inefficient time management, uneven student participation, memory burdens for student details. Through human-centered ideation, we distilled over thirty concepts into four widgets: student detail recaller, participation visualizer, slide summary overlay, and a visual timer, prototyped as a web-based simulation and Brilliant Labs AR Frames. In controlled mock-lecture sessions, AR assistance reduced pacing errors by 40%, increased recall accuracy by 29%, and raised unique student participation by 35%, with System Usability Scale ratings from “Good” to “Excellent.” These results suggest that context-aware AR overlays can significantly enhance lecture flow, personalization, and equity. We recommend native on-device deployment, integration with learning-management systems, and in-class trials to validate long-term educational impact and inform broader adoption
Calcium Connections: Developing a Reagent to Induce Cell Apoptosis and Assessing the State of Mental Health in Worcester’s Free Medical Programs
No full textCalcium is an important signaling molecule in both cellular models and neuropsychiatric function. CalBK is a fusion protein engineered to directly deliver calcium to the cytosol. It was hypothesized that CalBK could raise intracellular calcium levels and contribute to apoptosis. Initial findings from fluorometric and morphologic studies suggest that CalBK can function in these capacities. Abnormalities in neuronal calcium can be associated with psychological symptoms. An epidemiologic investigation and stress study were implemented to assess the state of mental health in free medical programs located in Worcester. The majority of patients surveyed reported moderate to high stress levels, indicating opportunities for expanding mental health resources