Department of Electrical & Computer Engineering Abstracts

Presenter(s): Clay Farley, Ethan Messier, Jaleel Khan, Andrew Rein

Showcase Advisor: Hany Elgala

Abstract: Manual resistor identification through color band reading is a time-intensive process susceptible to human error, particularly under suboptimal lighting or with aged components. These limitations present a recurring challenge in electronics laboratories and manufacturing environments where sorting accuracy directly impacts workflow efficiency. This project develops a computer vision system to automate resistor classification and address these shortcomings.

Because resistor identification depends on precise color pattern recognition rather than high-level semantic features, a CNN trained from scratch is better suited to this task than a pretrained ImageNet model. Images are preprocessed using color normalization and band segmentation to isolate task-relevant features prior to classification. Dataset augmentation—including rotation, brightness variation, and simulated component wear—is applied to improve robustness across real-world conditions.

Performance will be evaluated by benchmarking the CNN against traditional computer vision baselines using accuracy, precision, recall, and confusion matrix analysis, with final validation conducted through a working prototype sorting system.

Presenter(s): Toshin Ahnaf, Alexander Cutler, Thane Simera

Showcase Advisor: Dave Ardrey

Abstract: This is the precapstone project for the ECE program. Our project demonstrates the effects of various electrical components upon filtering noisy waveform signals produced by an electric guitar string and pickup sensor. During the demonstration one would simply pluck the guitar string and see the effect the filter has on a visual output. they could also change electrical component values and see the effect they have in difference to various trials.

Presenter(s): Brandon Morehouse, Colin Fields, Rohan Patrick

Showcase Advisor: Jonathan Muckell

Abstract: The purpose of this project was to create a non-contact water level monitoring systems for Global Foundries' chip fabrication machines. Before this solution, Global Foundries had no means of monitoring each fabrication machines coolant water level, and would have to shut down the machines to refill them as a result, resulting in lost production time. Using edge detection and image manipulation, this non-contact solution will be able to record water level data and transmit the said data to the Global Foundries server. This will result in an increase of production time for the company and a more efficient way for Global Foundries to monitor their multi-million dollar machines.

Presenter(s): Clay Farley, Daniel Levchenko, Mahib Rahman

Showcase Advisor: Mustafa Aksoy

Abstract: This capstone project, sponsored by the New York State Department of Transportation (NYSDOT), focuses on the development of a modernized software-based testing system for 2070 traffic signal controllers, which are routinely repaired and validated by NYSDOT. An existing testing setup became obsolete following the end-of-life of Windows 7, creating an operational need for a new, maintainable testing application and workflow. Student teams will pick up development of a partially completed tester application, extending and refining it to support single- and multi-controller testing scenarios, improve usability, and ensure long-term compatibility with current operating systems. The project emphasizes real-world software engineering, hardware–software integration, and test automation, with opportunities for continuity across multiple semesters and coordination with a NYSDOT summer intern.

Presenter(s): Kenneth Kowalski, Najani Johnson, Eric Dutan Sari

Showcase Advisor: Gary Saulnier

Abstract: This project presents the design and development of AccessAble, a lever-assisted rotational system created in collaboration with the Center for Disability Services Mail Fulfillment Center. Employees at the facility manually flip heavy stacks of paper during processing, which can lead to strain, fatigue, and inconsistent workflow efficiency. Our team designed a mechanical rotator mounted to a stable base that allows users to flip paper stacks using controlled rotational motion rather than lifting. The system integrates a reinforced support structure, flange-mounted rotation, and a lever-assisted mechanism to reduce required input force while maintaining stability under load. Safety and usability were prioritized through iterative prototyping, CAD modeling, drop testing, and structural analysis. The final design improves ergonomics, reduces physical strain, and promotes accessibility while remaining low-cost, durable, and practical for real-world implementation.

Presenter(s): Zaya Byambasambuu

Showcase Advisor: Hany Elgala

Abstract: Artificial intelligence–related courses are increasingly offered across multiple departments, but information about these offerings is often scattered across separate university webpages and catalogs. This project develops a prototype AI Curriculum Navigator for the University at Albany that aggregates AI-related courses from publicly available university sources into a single discovery interface. Instead of relying on database access or APIs, the system retrieves information directly from official webpages and reflects updates when those pages change, allowing the platform to remain aligned with existing university sources.

The project also incorporates surveys of students and faculty to evaluate interface preferences and inform the design of filtering and navigation features. By organizing distributed course information into a centralized interface, the prototype demonstrates a lightweight approach to improving course discovery and provides insight into how AI-related education is distributed across disciplines.

Presenter(s): Arge Gallo, Fariya Hossain, Anthony Papa

Showcase Advisor: Dave Ardrey

Abstract: The presentation will cover the prototype analog filtering board developed for the class project. The board includes adjustable, connectable analog filters that process a supplied audio signal. The filtered output is then played through a speaker and displayed on a screen alongside the original signal, allowing users to compare the two and observe how different filter types affect the audio.

Presenter(s): Glory Ajayi, Kendall Mariacher, Alex Canale

Showcase Advisor: Jonathan Muckell

Abstract: The Mail Sleeve Tray is an assistive electro-mechanical device designed to improve accessibility and efficiency within the Center for Disability Services Mailing Center. This project aims to support workers by reducing repetitive strain and minimizing the physical effort required to handle and separate mailing sleeves during daily operations.

Presenter(s): Johanna Brenord

Showcase Advisor: Aishwari Talhan

Abstract: Impaired balance increases the risk of falling, which is a major impediment to daily activity and rehabilitation. This study proposes a wearable, haptic jacket that blends tactile guidance and vision-based posture estimation to provide real-time balancing feedback. Actuators across the back offer directional haptic cues when postural deviations occur, while an AI-camera monitors the user's body position in reference to a specified stability region. Confirmation logic and filtering reduce jitter and prevent inaccurate feedback caused by transient motion or tracking loss. If the wearer veers off course or shows signs of unbalance, the jacket provides corrective vibration input. We plan to use this haptic jacket as part of a gamified VR-rehabilitation protocol that can improve posture and long-term stability while motivating repeated balance practice.

Presenter(s): Tyler Morrissey, Dillon O'Brien, Kemarli Thomas

Showcase Advisor: Dave Ardrey

Abstract: Our design project is an interactive educational board that demonstrates the ECE concept of digital logic gates and combinational logic circuits.  UAlbany’s Electrical and Computer Engineering department needs an interactive, cost-efficient logic gate demonstration system so that students can more effectively understand and apply Boolean logic and truth tables in circuit design.  Our system will allow users to arrange logic gates and adjust inputs in real time to create a desired truth table.  By immediately seeing how changing gates and inputs effect the output, students will strengthen their conceptual understanding and develop a clearer understanding of logic gate implementation.

Presenter(s): dylan zheng, Bowen Li, Lwin Hein, Jalloh Mohamed

Showcase Advisor: Hany Elgala

Abstract: This project develops an AI-based forecasting model for commercial and residential electricity usage. Using historical consumption data provided by National Grid (and other available utility or energy datasets), the model learns seasonal patterns and demand drivers to predict the amount of energy likely to be used in a given month. Key inputs include prior month and year usage trends, customer type (commercial vs. residential), estimated occupancy or number of people served, and geographic factors such as location and local climate indicators when available. The goal is to generate accurate monthly demand predictions that can support planning for grid operations, energy purchasing, and load management. Model performance is evaluated using standard error metrics on held-out historical periods, and results are compared across regions and customer classes to identify where forecasts are most reliable and where additional data features may improve accuracy.

Presenter(s): AUNTORA ROY CHOWDHURY

Showcase Advisor: Dola Saha

Abstract: ADS-B stands for Automatic Dependent Surveillance–Broadcast. It is a modern surveillance technology that aircraft use to automatically broadcast their position information, such as latitude, longitude, speed, flight ID, etc., every second. It is used by the air traffic control system to monitor air traffic and effectively manage the sky. ADS-B messages are transmitted at 1090 MHz, and the signals can be received and decoded by anyone using suitable hardware and signal processing methods. This poster presents a practical implementation of ADS-B signal reception using a software-defined radio (SDR) receiver, and captured radio signals are processed and decoded in MATLAB to extract aircraft information from the raw transmissions. This work demonstrates the end-to-end pipeline of signal acquisition, demodulation, and message decoding, illustrating how low-cost SDR platforms can be used to observe and analyze real-world air traffic broadcasts.

Presenter(s): Joshua Muhammad, Marianna Karagiannis, Faiyaz Sajjad

Showcase Advisor: Hany Elgala

Abstract: Li-Fi (Light Fidelity) is a wireless communication technology that transmits high-speed data using visible light from LEDs, offering speeds up to 100 times faster than Wi-Fi in lab tests while completely avoiding radio-frequency interference. This makes it ideal for sensitive environments like hospitals, airplanes, and factories. Despite its advantages, Li-Fi adoption is limited by expensive, proprietary systems that hinder research and practical use. Our specific focus is designing a unidirectional Li-Fi prototype by using processing and analog components to improve speed, reliability, and accuracy. We will integrate a W5500 Ethernet interface to provide direct, standardized network connectivity, a 600 MHz Teensy 4.0 microcontroller to serialize and deserialize data streams, driving a direct MOSFET transmitter to process incoming optical signals through an optimized Transimpedance Amplifier (TIA) and a high-speed LMV7219M5 comparator. The result is an open-platform, Ethernet-connected Li-Fi system under $100 per module, ready for real-world testing.

Presenter(s): Powell Burke, Tierney Lawrence

Showcase Advisor: Bariscan Yonel

Abstract: EASyExam is an end-to-end deep learning system that automates ultrasound video classification through an integrated graphical user interface. Built with PyTorch and accelerated with GPU support, the system processes ultrasound videos and delivers real-time classification results within an intuitive drag-and-drop interface. By combining advanced AI with practical usability, EASyExam transforms complex medical imaging analysis into an accessible and deployable clinical tool.

Presenter(s): Ethan Messier, Rachel Gay, Jacob Mohammed, Aaraiz Masood

Showcase Advisor: Dave Ardrey

Abstract: This project uses a Blackjack-style game as a hands-on educational tool to demonstrate how Finite State Machines (FSMs) work. The system moves through stages such as betting, dealing cards, player actions, dealer actions, and determining the winner. Each stage represents a different state, helping visualize how an FSM transitions between steps to control a system. Users interact with the system using physical buttons, while LEDs and a 20x4 LCD display show the current state and game information. The goal of this project is to create a simple and interactive way to help students understand how FSMs control the behavior of digital and embedded systems. By using a familiar game structure, the project provides a clear demonstration of how step-based logic can be used to design and control real electronic systems.

Presenter(s): Kaijun Cao, Zaid Akari, Brandon Morehouse

Showcase Advisor: Hany Elgala

Abstract: This project investigates machine learning methods for detecting cardiac abnormalities from electrocardiogram (ECG) signals under realistic noise conditions. Using the MIT-BIH Arrhythmia Database, ECG recordings are segmented into heartbeat windows and preprocessed through filtering and normalization. The problem is formulated as a supervised classification task to distinguish normal and abnormal heartbeats. Two modeling approaches are compared: a classical machine learning pipeline using manual feature extraction, Principal Component Analysis (PCA), and a Support Vector Machine (SVM); and a neural network approach using a one-dimensional Convolutional Neural Network (1D-CNN), which applies convolutional filters along the temporal axis of ECG signals to automatically learn local waveform features before classification through fully connected layers. Model performance is evaluated using accuracy, precision, recall, F1-score, and confusion matrix analysis, with emphasis on minimizing false negatives. The study highlights trade-offs between interpretability and automatic feature learning, while assessing robustness and generalization in noisy physiological signal environments.

Presenter(s): Colin Fields, Brandon Morehouse, Rohan Patrick

Showcase Advisor: Jonathan Muckell

Abstract: Accurate monitoring of coolant reservoir levels in photolithography equipment is critical for maintaining process stability and preventing unplanned downtime in semiconductor manufacturing. This project, conducted in collaboration with GlobalFoundries, presents a non contact fluid level measurement system using an embedded computer vision approach. A Raspberry Pi camera captures real time images of the reservoir sight glass, and a custom image processing pipeline detects the region of interest, extracts the waterline using edge detection and thresholding techniques, and converts the result into a percentage based level measurement. The system logs validated readings to a CSV file for traceability and future integration with facility monitoring infrastructure. This solution eliminates the need for intrusive sensors, reducing contamination risk and installation complexity while providing a low-cost and scalable alternative for industrial deployment. Experimental testing demonstrates reliable waterline detection under varying lighting and alignment conditions, supporting its feasibility for use in semiconductor manufacturing environments.

Presenter(s): Toluwani Alao

Showcase Advisor: Aishwari Talhan

Abstract: Integrating a sense of “feel” into Video calling/conferences with Al-Driven Gesture Recognition and Haptic Feedback enacted by haptic motors.

Presenter(s): Zaid Akari, Meghan Herbert, Kai Perez

Showcase Advisor: Jonathan Muckell

Abstract: The aim of this project is to design and build a TCD (Thermal Conductivity Detector). In this design we would like to incorporate several small TCD devices, that could work in an array to measure an area with data. This should aim to be scalable and reproducible at a small cost, without as much concentration into accuracy, but instead the actual area to get a reading in which it suspects an area with a higher than normal concentration of Helium than expected, indicating a leak. This design relies on the principles that the actual thermoconductivity of Helium would present a lower value making the resistance higher or lower based on the type of thermistor used.

Presenter(s): Kalonji Samuel

Showcase Advisor: Hany Elgala

Abstract: As generative AI reshapes higher education, faculty are frequently caught between sweeping institutional mandates - like FERPA and the NY Empire AI initiative - and the nuanced realities of teaching. Blanket policies often fail to serve specific learning objectives. To address this friction, this project introduces a dynamic web application that empowers educators with true pedagogical flexibility.

Moving beyond rigid, course-level addendums, this tool enables granular, assignment-level governance. Faculty can upload their course schedules and assign specific AI usage parameters and tiers - ranging from maximally restrictive to fully encouraging - directly to individual tasks and projects. The engine then seamlessly integrates these customized guidelines natively into the syllabus. By embedding acceptable use cases, prompt-logging mandates, and equity-based alternatives directly into the coursework, this framework bridges the gap between top-down data governance and grassroots academic freedom, providing a scalable model for modern university AI policy administration.

Presenter(s): Sanad Sahawneh, Brian Patterson, Benjamin Bellenchia

Showcase Advisor: Jonathan Muckell

Abstract: The Green Mountain Semiconductor (GMS) project is a senior capstone initiative focused on the design, verification, and testing of radiation-tolerant AI hardware structures intended for aerospace applications. Our team is developing and validating key on-chip subsystems—including scan chains, SRAM test structures, ring oscillators, and inverter chains—to ensure reliable operation in harsh environments such as space. Emphasis is placed on structured system testing, traceability, and hardware-software co-verification prior to silicon validation. Through a disciplined engineering approach, the project integrates simulation, measurable verification procedures, and structured documentation to ensure each subsystem meets functional and reliability requirements. By combining pre-silicon modeling with systematic hardware validation strategies, the GMS initiative aims to enhance confidence in semiconductor performance under extreme environmental conditions and contribute to the advancement of resilient, mission-critical AI architectures for next-generation aerospace systems.

Presenter(s): Christian Francisco, Rick Gue, Kc Kc

Showcase Advisor: Dave Ardrey

Abstract: This project presents a human powered electromechanical generator designed to be a interactive educational tool to demonstrate Faraday's Law of Induction and Lenz's Law. By turning a hand cranked gear train, users drive a magnetic rotor over stator coils to produce alternating current. As the user experiences the physical mechanical resistance dictated by Lenz's Law; a data recording system captures the resulting electrical output. This telemetry (voltage, current, and magnetic flux) is streamed to a live display, providing the user with a tangible, real-time visualization that directly correlates the user's physical effort with electromagnetic principles.

Presenter(s): Jason Armoo

Showcase Advisor: Aishwari Talhan

Abstract: This project is based on designing a wearable haptic navigation system that helps a person navigate towards a target device using Bluetooth signals. A belt using a Raspberry Pi measures the signal strength of a Raspberry Pi Pico beacon and provides directional vibrations based on the direction of the target device. The system also makes use of AI-based vision sensors and ultrasonic sonar sensors for obstacle detection during navigation. This helps in converting various signals into haptic feedback, allowing users to navigate through environments easily.

Presenter(s): Jaleel Khan, Tyler Sljukic, Andrew Rein, Amari Olivacce

Showcase Advisor: Jonathan Muckell

Abstract: It is a capstone project that involves the design and development of a custom daughter board that interfaces with the BeagleBone single-board computer. Our goal is to develop and fabricate a PCB, develop and validate low-level C drivers for the board's I/O.

Presenter(s): Justin Ault, Elizabeth Ortiz, Johanna Brenord, Michael Cardarelli

Showcase Advisor: Dave Ardrey

Abstract: This presentation will show our finished project and the process of our design and implementation. The project focuses on a specific Electrical and Computer Engineering concept and how you can teach ECE students this topic. Our topic that we are teaching is Combinational Logic Gates through a combinational logic puzzle.

Presenter(s): Colleen McHugh, Sadman Khan, Raffaella Bongiovanni, Zane Schulties

Showcase Advisor: Hany Elgala

Abstract: Liver cirrhosis is a progressive and potentially life-threatening condition in which accurate staging is essential for guiding treatment decisions, monitoring disease progression, and estimating patient prognosis. However, distinguishing between cirrhosis stages can be challenging because clinical indicators often overlap, and no single measurement fully captures disease severity. This complexity makes traditional classification methods prone to variability and subjectivity. By leveraging patterns across multiple clinical features simultaneously, machine learning offers a data-driven approach to support more consistent and reliable classification of liver disease severity. The dataset being used is from the Mayo Clinic; an open-source set with over two gigabytes of data, with each data point having 19 features. By using machine learning models such as Decision Trees, Random Forests, and Neural Networks, this project aims to demonstrate how data-driven approaches can support healthcare workers in making more efficient diagnostic decisions.

Presenter(s): Colin Fields, Tierney Lawrence

Showcase Advisor: Hany Elgala

Abstract: Lithium-ion battery health prediction is essential for improving reliability, safety, and lifecycle management in modern electronic and energy storage systems. This project presents a data driven approach for modeling battery capacity degradation using machine learning techniques. A publicly available battery aging dataset is processed to extract relevant features related to charge / discharge behavior and cycle life. Multiple models, including classical machine learning algorithms and neural networks, are trained to predict remaining capacity as a function of usage. The performance of these models is evaluated and compared using standard regression metrics to determine their accuracy and generalization capability. The results demonstrate the effectiveness of machine learning for capturing nonlinear degradation trends and provide insight into the trade offs between model complexity, prediction accuracy, and computational cost. This work highlights the potential of data driven methods for battery health monitoring and predictive maintenance applications.

Presenter(s): Powell Burke

Showcase Advisor: Hany Elgala

Abstract: Machine learning is increasingly being applied to echocardiographic video analysis for tasks such as cardiac view classification, cardiac feature detection, and image labeling. Video of the heart poses a challenge that deep networks struggle with: temporal rate variance. Hearts which look physiologically similar produce different signals for networks to learn from when beating at different rates, making learning more difficult. Even when models can handle this rate variance to some extent, they require more training data and compute resources to do so, with the former being highly constrained in many medical settings. To address this, this project attempts to automate heart rate estimation from echocardiographic video clips. This allows frames to be sampled at consistent cardiac cycle intervals, normalizing temporal rate and preventing temporal aliasing.

Presenter(s): Najani Johnson, Meghan Herbert

Showcase Advisor: Hany Elgala

Abstract: Semiconductor manufacturing relies on wafer inspection to identify defect patterns that can reduce device yield. These defects often appear as spatial patterns on wafer maps, which are traditionally analyzed through manual inspection or rule-based methods. In this project, we propose using machine learning techniques to automatically classify wafer defect patterns from wafer map images. Using the WM-811K wafer map dataset, we will treat the problem as a multi-class image classification task. Two approaches will be explored: classical machine learning models using handcrafted spatial features and a convolutional neural network (CNN) that learns features directly from the wafer images. Model performance will be evaluated using metrics such as accuracy, F1-score, and confusion matrices. This project aims to compare the effectiveness of classical and deep learning methods for identifying defect patterns in semiconductor manufacturing data.

Presenter(s): Selim Bora GENCOGLU

Showcase Advisor: Zi Yang

Abstract: Standard optimization methods like Gradient Descent often struggle in high-dimensional or constrained spaces where Euclidean geometry is not the natural fit. This project explores Mirror Descent Algorithm (MDA), a powerful generalization that uses Bregman Divergence to adapt to a problem’s intrinsic geometry. By performing updates in a "dual space" via a mirror map, MDA provides a more flexible and efficient framework for non-Euclidean optimization. We analyze the theoretical foundations of the algorithm and demonstrate its practical superiority in convex optimization.

Presenter(s): Finn Bastian, Tolu Alao, Safwan Shikder

Showcase Advisor: Hany Elgala

Abstract: With the rapid expansion of digital music streaming platforms such as Spotify, 
Apple Music, and Pandora, there is a growing need for automated systems that can accurately classify millions of new audio tracks into musical genres. Manual genre labeling is time consuming, subjective, and does not scale to modern music libraries. This project focuses on the engineering challenge of extracting meaningful information from raw audio signals and using machine learning techniques to automatically classify music genres. The goal of this project is to explore and compare traditional signal-processing-based machine learning methods with modern neural network approaches, highlighting their strengths, limitations, and performance trade-offs.

Presenter(s): Colleen McHugh, Alondra Cruz-Delgado, Zane Schulties, Amanda Astudillo

Showcase Advisor: Mohammed Agamy

Abstract: This project involves designing, implementing, and testing innovative methods for powering embedded systems without relying on traditional power sources. Students explore alternative energy harvesting techniques, such as solar, vibration (haptic) energy, and heat (Peltier effect), to power an ultra-low power embedded microcontroller. The system will handle a range of various scenarios and be designed with careful consideration of stakeholder design constraints. The project's ability to integrate these energy sources effectively, ensuring reliable operation even under variable environmental conditions is key. Students will evaluate the long-term sustainability and efficiency of their solutions, aiming to meet or exceed the standards expected in real-world applications.

Presenter(s): Hasti Ataei Zolfaghari

Showcase Advisor: Dola Saha

Abstract: Recent advances in machine learning have created new opportunities for improving wireless communication systems. Traditional communication algorithms are typically designed using mathematical channel models, which often fail to capture the complexity of real-world wireless environments. In this project, we explore the idea of training neural networks directly over the air, where learning occurs through real wireless transmissions instead of relying only on simulated channels.

In this framework, a neural network–based transmitter and receiver interact with the physical wireless channel during training. This allows the system to automatically adapt to real-world effects such as noise, interference, fading, and hardware imperfections. By including the wireless channel inside the learning process, the system can discover signal representations that improve communication reliability and robustness.

This work presents a conceptual framework and an initial implementation of over-the-air neural network training, illustrating how machine learning can enable more adaptive and intelligent wireless communication systems.

Presenter(s): Justin Kendall, Renee La Londe

Showcase Advisor: Mohammed Agamy

Abstract: There are many spaces in industrial environments where the use of traditional wall power sources would be impractical or unachievable. Using energy harvesting techniques provides an alternative power solution in these harsh environments and ensures devices can remain functional in the event of a power malfunction or failure. Our solution uses Peltier modules to harvest energy from heat and uses that energy to run a low power microcontroller capable of running a temperature sensor and SD card data logger.

Presenter(s): Sanad Sahawneh, Brian Allen, Mahib Rahman

Showcase Advisor: Hany Elgala

Abstract: This project explores machine learning-based prediction of marathon finish times using data from the 2023 Boston Marathon dataset available through the SCORE Network Sports Data Repository. The dataset contains performance records for 26,598 runners. We formulate the task as a supervised regression problem, where the goal is to predict a runner’s net finish time in seconds using demographic and performance features such as age range, gender, and intermediate split times. After cleaning and preprocessing the data, including handling missing values and encoding categorical variables, we implement two predictive models: a linear regression model and a feedforward neural network (multi-layer perceptron) regression model. Model performance is evaluated using Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). By comparing these approaches, we assess whether the neural network’s ability to capture nonlinear relationships improves predictive accuracy. This work demonstrates how machine learning can be applied to athletic performance analytics.

Presenter(s): Saugat Shah, Nze Nwabara, Christian Zhinin-Avila

Showcase Advisor: Dave Ardrey

Abstract: Project SkyLog(ic) is a junior design project that demonstrates a closed-loop helicopter lift system inspired by a drop-tower structure. The system uses an Arduino, an ultrasonic distance sensor, and a pulley-driven mechanical assembly to control the vertical motion of a helicopter with spinning blades. A PID control algorithm continuously compares the measured distance to a desired setpoint and adjusts motor output to move the helicopter up or down smoothly and accurately. This project integrates mechanical design, electronics, sensing, and control theory into one working prototype. It highlights how feedback systems can improve stability, positioning, and responsiveness in electromechanical applications and provides a practical demonstration of embedded control concepts.

Presenter(s): Sanad Sahawneh, Brian Allen

Showcase Advisor: Jonathan Muckell

Abstract: This capstone project develops a low-cost RF exposure monitoring system enhanced with TinyML for real-time classification of environmental RF levels. Building on existing RF sensing hardware, the system measures peak and RMS voltages proportional to incident RF power using a 3.3 V microcontroller and ADC. Data is processed locally and transmitted via LoRa for logging and analysis. A supervised decision tree model is trained on collected sensor data to classify exposure levels as Low, Medium, or High. The trained model is converted into a lightweight format and deployed directly on the embedded hardware, enabling on-device inference without cloud dependence. This reduces latency, power consumption, and communication overhead. System validation includes calibration, dataset collection across multiple environments, model evaluation, and embedded deployment testing. The project demonstrates the feasibility of integrating RF sensing hardware with embedded machine learning for scalable, energy-efficient environmental monitoring applications.

Presenter(s): Nolan Callaghan, Kelvin Biney, Bowen Li, Sadman Khan

Showcase Advisor: James Moulic

Abstract: This project develops a scalable Shared Silicon Framework (SSF) that enables multiple independent Verilog modules to be integrated within a single ASIC architecture. The framework allows up to four modules to share input and output resources through controlled module selection and output multiplexing while maintaining modularity and preventing signal conflicts. To improve usability and testing flexibility, a UART-based communication interface is implemented, enabling software-driven interaction with the hardware through a PC terminal. The system is first validated on the Nexys A7-100T FPGA using Vivado before transitioning into the ASIC design flow using Cadence tools and the SkyWater 130nm PDK. By enabling multiple student designs to share a single chip architecture, the framework supports cost-efficient fabrication and provides students with hands-on experience in digital design, system integration, and ASIC verification workflows.

Presenter(s): Sachintha Kavishan Jayarathne

Showcase Advisor: Seetal Potluri

Abstract: Edge neural processing units (NPUs) function within a trusted execution environment (TEE) that relies heavily on memory encryption to secure off-chip (DRAM) traffic using hardware AES engines. However, off-chip traffic to the accelerator on edge NPUs substantially increases energy consumption by a factor of ten. We make two important observations regarding On-Device AI Inference: (a) Feature tensor encryption incurs significantly more overhead than filter tensor encryption in AI models, making it unacceptable for resource-constrained edge devices. To address this, we propose “Selective Encryption,” which reduces feature encryption overhead by targeting intermediate layers and their output feature maps (OFMAPs); (b) Encrypting the intermediate OFMAPs of certain layers based on their sensitivity and risk of exposure to reverse engineering attacks enables us to maintain security while minimizing unnecessary encryption overhead.

Presenter(s): Finn Bastian, Toluwani Alao, Safwan Shikder, Alex McDougall

Showcase Advisor: Jonathan Muckell

Abstract: This project will create a modular ASIC framework capable of supporting up to four independently developed Verilog modules. At runtime, a control signal will select which module is active, allowing the modules to share I/O ports without interference. The final design must satisfy all standard fabrication and verification requirements, including DRC (Design Rule Check): ensuring compliance with process-specific layout constraints, LVS (Layout vs. Schematic): verifying that the physical layout matches the intended logic, and additional verification steps required by the foundry. By consolidating multiple projects into a single shared tape-out, this approach will reduce fabrication costs while also preparing students with workforce-ready skills in modern semiconductor design and verification. The framework will be well-documented, modular, and designed for scalability to support future teams with minimal interdependencies.

Presenter(s): Eric Crespo, Isabelle Keovongxay, David Green

Showcase Advisor: Nathan Dahlin

Abstract: Solar energy deployment in New York has accelerated rapidly in recent years. As distributed photovoltaic capacity continues to grow, accurate behind-the-meter solar generation forecasts have become essential for reducing operational costs, improving grid reliability, and enabling efficient market participation. However, solar production is inherently intermittent. Variability driven by cloud cover, precipitation, and other weather phenomena makes reliable day-ahead forecasting a persistent challenge. 
In this project, we employ a novel machine learning approach based on the TabPFN tabular foundation model to generate day-ahead forecasts across NYISO load zones. TabPFN is pre-trained on a large and diverse collection of synthetic datasets and tasks and can achieve strong predictive skill with limited training data through in-context learning. For model development and back-testing, we use meteorological predictors derived from the ERA5 reanalysis dataset. For real-time forecasting, we replace these inputs with corresponding fields from the Global Forecast System (GFS), enabling seamless operational deployment.

Presenter(s): Ryan Patterson, Tyler Bissonnette, Alice Mazy, Jadon Etuale

Showcase Advisor: Dola Saha

Abstract: Noise pollution is problematic and difficult to quantify for the general population in both urban and rural areas. It leads to numerous health conditions which can be unavoidable; specifically for people under financial domicile constraints, i.e., people who cannot easily relocate. To help mitigate this issue, meaningful data must be gathered and presented for free to the public such that the public may be informed either for future planning or community outreach.

Presenter(s): Prakash Kota

Showcase Advisor: Hany Elgala

Abstract: Meetings generate valuable discussions, decisions, and action items, yet much of this information is often lost or difficult to retrieve afterward. This poster presents the UAlbany Meeting Agent (UMA), an AI-based system that converts meeting transcripts from platforms such as Microsoft Teams and Zoom into structured meeting intelligence. UMA operates as an ex post agent: users upload transcript files (.vtt), and the system applies natural language processing to generate targeted outputs including executive summaries, action items, decision logs, and open questions or risks. The architecture includes transcript parsing, normalization, context construction, and a single-mode AI reasoning layer that produces concise governance-ready reports. By transforming raw transcripts into structured insights, UMA aims to improve meeting transparency, institutional memory, and follow-through without requiring changes to existing meeting platforms. The prototype demonstrates how lightweight AI agents can augment collaboration workflows across academic and organizational settings.