CSE 2025 Spring Departmental Demo Day

(Blockchain Think Lab)

Blind Match

• rohanshr
• mohantej

We propose a library in Golang to for parallelized homomorphic cosine similarity operations. To demonstrate our implementation we also present Blind Match. A Fuzzy Private set intersection protocol to efficiently compare encrypted entity data using TF-IDF vectors while preserving privacy. To reduce the computational cost of homomorphic encryption, we exploit the sparsity of n-gram TF-IDF vectors and apply com- pression techniques that maintain cosine similarity, such as Fourier and Orthonormal Wavelet Transforms. Our project leverages the CKKS encryption scheme for secure dot product computations and is developed as an open-source Go Implementation

(CSE 302/303/402)

Materials Microstructure

• laibahah
• joseph55
• donghwis
• kem44

Over the ages, new materials have driven innovation and shaped our civilization. The names of the main prehistoric phases of human history, e.g., the Stone Age, Bronze Age, Iron Age, etc., are the testimony to this statement. The progress has been paralleled by a better understanding of the relationship between (micro)structure and material properties that led to the plethora of materials (e.g., reinforced concrete, LEDs, graphene, organic semiconductors). In recent years, progress in materials research has been fueled by machine learning (ML) and artificial intelligence (AI). To streamline the transition, the information about the materials' structure needs to be converted into an ML-readable format. GraSPI is a software developed at UB that featurizes the micrographs into an array of physically meaningful descriptors that can be directly used in ML pipeline. In this project, students will enhance an existing GraSPI project. GraSPI uses a graph/network as a data structure to efficiently calculate descriptors from materials micrographs. The project's current version is coded in C/C++ using the boost library. This project aims to translate the current implementation to be python-native (e.g., using NetworkX) or use a boost.Python library. Part of the project will be: - to identify the best strategy for the needed translation (research solution, plan and execute the basic tests with classic graph-based algorithms, and make the suggestion); - plan the translation between two packages - implement the Python packages (core functionality, documentation, example notebooks)

Coriolis-lite (SMARTEn)

• Paridhi Desai
• Purushotham Sirasapalli
• Chris Chen
• Vatsal Labh
• Troy Richardson

The Coriolis-lite project implements efficient querying of an FM-Index using the SDSL library. It allows users to search a precomputed index for longest matching substrings based on provided queries.

ML Enhanced Indoor Wireless Localization and Navigation

• Joey Normil
• Chongze Peng
• Harrison Moore
• Yufeng Chen
• Yang Hu
• Sai Roshan Ayyalasomayajula

This is a navigational Android app designed to help users find specific locations inside buildings (specifically davis hall), such as classrooms, stairs, elevators, and bathrooms. It enhances indoor navigation by using data collected from the app to improve the accuracy of a user’s location on a map through machine learning.

Find a Mechanic

• Yudian Ke
• Esteban Perez
• Alex McHugh
• Weitian Wu

Find a mechanic of FAM for short is a service based webpage and app with the goal of becoming a doordash and/or instacart of trades and services. This software has 3 core portions. A marketplace webpage for a shop to administer and delegate time, work, and resources internal to customer issues and communicate and do the entire workflow process with customers. The other two portions are an app that has two uses. 1st use is for a worker of the shop to have work delegated to them from the shop webpage that allows them to update on progress of a vehicle and a user facing app that allows them to find a shop of their desired niche by reviews. The user can interact with the shop from start to finish from the app. As of right now FAM is nearly finished as its most basic core tenets.

DevU

• diegocab
• rsahu
• tasneemz
• sameepko

DevU is an automated software-grading platform being developed at the University at Buffalo. DevU aims to be incredibly extensible, allowing professors to add any functionality they desire without reaching a dead end. It will eventually replace Autolab and other services.

Choreographic Programming (Pirouette)

• Alex Doyoon Kim
• Claire Huyck
• Ethan Canton
• Julia Montuori
• Yining Zhang

We are the undergraduate team for Pirouette, a new language for choreographic programming. Choreographic programming aims to simplify distributed systems / concurrent systems programming by allowing a programmer to specify the system in its entirety - participants, their computations, and communication patterns. From this high level choreographic program, we can synthesize concrete implementations for each node in the distributed system or each thread in the concurrent system.

Encrypted Storage

• Benjamin Anderson
• Aaron Radliff
• Liam Cassidy
• Weiguo Wu

A system for encrypted storage where the decryption key is split among multiple devices and compromise of a device with a partial key does not lead to disclosure of the encrypted content.

AI Energy Optimization

• JakeNathanson
• LayneXia
• CristianRivera
• AhmedBouhraoua
• RobMurray rmurray5
• AsifImran

We aim to design and develop a tool that will analyze popular ML frameworks like pytorch, tensorflow, dusk, pandas, etc and will be able to calculate gather data at the method level to estimate energy consumption and carbon emission for big data workloads leveraging those frameworks. Next, we would like to us ML itself in the tool to suggest how making changes to the structure of the code can improve the energy usage for the same workload. Overall, we are planning to design an automated refactoring tool that will refactor from the perspective of energy optimization and not only code quality. The team needs to have knowledge on software engineering processes like requirement engineering, software quality assurance and testing, python language, abstract syntax trees (AST) and software documentation. We can divide the project in the following phases: Phase 1: Identify the attributes in ML frameworks which contribute to energy consumption. These can be specific methods, classes, etc which consumes significant energy during execution. Phase 2: Build a software that will automatically execute workloads on the ML frameworks and collect energy data at the GPU level. Gather this data ta the method level. Phase 3: Use ML to train the software to suggest refactoring of ML frameworks to optimize energy usage. Phase 4: Execute automatic refactoring of the codes based on the feedback of phase 3.

(CSE 370: Applied HCI)

Melting Pot

• Vaishnavi Lokhande
• Joshua Castillo
• Yessica
• Justin N

To simplify others cooking experiences with others Share your own experiences with others, the negatives, the positives Be able to cater to to each user’s specific needs and preferences Allergies Diets Cooking skill level Make users with related interests interact with one another

Bubble Brain

• Rances
• Cayden
• Jacob
• Tariq
• Akib

A fun studying App

(CSE 4/546 - Reinforcement Learning)

Real-Time Adaptive News Recommendation

• lalasamy
• vgawale

The project intends to use an appropriate reinforcement learning (RL) technique to dynamically recommend the news items from instant user interactions so as to effectively solve real-time changes in user's preferences for a particular topic.

NegotiableAI: Multi-Agent Reinforcement Learning for Real-Time Buyer–Seller Negotiations

• Sada Kakarla
• Shivansh Gupta
• Aditi Sinha

This project automates the negotiation process between buyer and seller agents using a Multi-Agent Reinforcement Learning (MARL) framework. This project helps to meed the need for an intelligent solution that can simulate human-like bargaining behavior, adjust offers in real time, and maximize mutual benefit between negotiating parties without manual intervention. Each agent is trained using Proximal Policy Optimization (PPO), leveraging self-play to dynamically evolve negotiation strategies over alternating offers and counteroffers. By simulating real-world negotiation dynamics, the system aims to build adaptive, scalable, and efficient autonomous negotiators capable of handling complex and evolving deal-making environments. Negotiations are formulated as a sequential decision-making process with bounded action spaces, where agents alternately propose and counter-propose offers.The use of clipped surrogate objectives in PPO ensures stable policy updates under the non-stationary dynamics of multi-agent learning. The system is designed to enable adaptive and rational negotiation strategies applicable to domains such as e-commerce pricing, automated trading, B2B deal-making, and resource allocation systems.

Autonomous Multi-Drone Package Delivery System for Efficient Warehouse to Customer Logistics Using Multi-Agent Reinforcement Learning.

• Sarthak Sharma
• Annoy Dey
• Md Arman Islam

Two drones (left drone and right drone) are used for delivering packages from a warehouse to destinations. The environment is a 2-dimensional grid of size 10 x 5. Warehouse is in center of grid's right half. Package destination can be any random location in the grid. Ideal behaviors: If destination is in grid's right half, then the right drone delivers the package, and the left drone stays idle. If the destination is in grid's left half, then the right drone transfers the package to the left drone, and then the left drone delivers the package. We implement a suitable reward structure for to encourage the above ideal behaviors. Drones' battery levels and chargers may be included if time permits for the additional complexity that this brings.

Red Light, Green Light; RL for Traffic lights

• Daniel Viola
• Erin Gregoire

we are using a variety of agents to train small traffic lights, as well as double deep q learning as well as tentatively PPO to train the 25 traffic lights in the block surrounding UB's Medical campus to optimize traffic flows!

Intelligent Accident Risk Prediction and Resource Allocation

• Sahil Sawant
• Meghna Trichur Saravana Shekhar
• Fagun Patel

This project explores the use of Reinforcement Learning (RL) to predict high-risk traffic accident zones and optimize emergency resource allocation using real-world accident data from Buffalo city. In this phase of the project, we worked on developing a reinforcement learning (RL) model aimed at improving how emergency resources (like ambulances, tow trucks, and police units) are allocated across different zones in the Buffalo area.

RL Team 9 - RKP vs RFID

• Rahul Ekambaram
• Pavithran Gnanasekaran
• Kanisha Raja

Our project aims to tackle the RFID reader collision in crowded IoT settings like smart warehouses using multiple reinforcement learning techniques. We built a 10x10 grid simulation where multiple readers pick frequency and time slots while avoiding interference, signal jammers, and shifting traffic. The setup is designed keeping in mind the real-world challenges with added randomness and traffic variations. We completed testing on five RL strategies: the original HiQ Q-learning, our custom Q-learning, SARSA, DQN, and Double DQN. Each agent is made to learn to minimize collisions and maximize successful tag reads through smart frequency-slot selection.

(CSE 442: Software Engineering)

O(no)

• Songzhu Li
• Daniel Gulvin
• Jeffrey Huynh
• Johnathan Tang
• Peter Vaiciulis

A WebApp that helps people find recipes based on their dietary needs, and points them to where they can buy the ingredients. The vision is to empower young people—particularly those in middle or lower income brackets—to make healthier food choices without breaking the bank. Through an intuitive webapp, users can create weekly meal plans tailored to their dietary preferences and lifestyle goals, explore culturally diverse recipe recommendations, and track nutritional intake for balanced meals. By leveraging real-time inventory data from local grocery stores, the app also shows users where to buy ingredients affordably and conveniently. Ultimately, the project aims to simplify the journey toward healthier eating by providing personalized guidance, cost-saving tips, and actionable insights—making nutritious living accessible to everyone.

ALLBREAKSNOGAS

• orkop
• elishaam
• nxfergus
• aidenfer
• chuyck

Our team has developed a 2D Buffalo themed game, which incorporates a retro 8-bit style with Buffalo scenery and characters to deliver a fun and special gaming experience. We have four single-player levels, alongside an online leaderboard to track and compare stats across these levels. Furthermore, we also have CSE442's first working real-time multiplayer system, allowing two users to join a lobby system and play a multiplayer level against one another. We also have a points shop and badge system, where players can use the coins they have collected throughout the single-player levels to purchase unique collectible badges.

Team Rizzipe

• Zowraiz Qureshi
• Wyatt Premore
• Oscar Li
• Chongze Peng
• Chang Luo

Created a fully functioning web application that allows users to create, share, and discover recipes, as well as plan meals throughout the month, and critique recipes posted by other users.

FlowVesta

• sakibhas
• colinwon
• tlee47
• gvidojau

Stock tracker social media

Next gen uno

• rances
• kurian
• Kendall
• Haoyan
• Cody

a multiplayer Uno game with Gambling

WAYFARER

• orniepay
• ezhang8
• kpang
• mnshabbi
• mnshabbi

WAYFARER, a 2D platformer that breaks the mold of grindy, repetitive games and delivers the adventure you’ve always wanted. Traverse vibrant, distinct biomes packed with secrets, where every jump and gadget unlocks new ways to explore. Discover powerful gadgets and special collectibles as you unravel an immersive story that keeps you hooked from start to finish. With unique powers, clever level design, and a world that rewards curiosity, WAYFARER is everything you love about platformers without the clutter or frustration.

(CSE4/555: --Pattern Recognition)

AKN Squad

• Aishwarya Virigineni
• Nithya Kaandru
• Kanisha Raja

This project improves clinical trust in Bayesian medical image segmentation by aligning uncertainty with actual errors. AvU loss is applied to segmentation tasks, a rarely explored direction in medical AI. Simulated rater disagreement approximates aleatoric uncertainty without requiring costly multi-rater datasets. Both epistemic and aleatoric uncertainties are unified into a simple, demo-friendly framework. The resulting uncertainty maps are designed to hold real clinical significance, beyond being mere visual artifacts.

Facial Emotion Recognition using Deep Learning

• lalasamy
• vgawale
• christjo

Facial Emotion Recognition using Deep Learning is a computer vision project that aims to identify human emotions from facial expressions in images/videos using deep learning techniques.

Lumon Industries

• Atul Pandey
• Atharva Prabhu
• Kalash Thakur

Satellite Image Analysis for Land Cover Classification

SoundSense

• Gagan Vadlamudi
• Patanjali Uppugandla
• Muni Sai Kannemmagari

This project aims to build a system that performs Speaker Identification and Speaker Emotion Recognition from audio data. The goal is to accurately identify who is speaking and determine their emotional state (e.g., happy, sad, angry) based solely on their voice characteristics. The system uses audio preprocessing, feature extraction, and machine learning or deep learning models to classify speakers and emotions. Applications include voice-based authentication, emotion-aware virtual assistants, and human-computer interaction.

Heart Disease Analysis

• bhuvanth
• rkulothu
• manojmah

The project will predict at what age, one can have their first Heart Stroke based on Current Health Parameters. Our Aim is to forecast the future health parameter based on current state and then apply classification model to identify the chances of heart stroke.

Leveraging Pattern Recognition to Detect Revenue Leakage in Healthcare

• Rahul Ekambaram
• Rishab Darshan Shylandra

This project involves in the development of an AI-based system designed to detect and reduce "revenue leakage" in the healthcare sector. While leveraging datasets related to real-world claims and operational datasets, the solution is mainly used while applying ML and graph-based techniques to understand and uncover patterns of fraud and process inefficiencies. The idea to integrate explainable AI with custom-built dashboards is proposed to help in financial decision-making ability/analysis.

Demand Forecasting For Grocery sales using Transformers

• csomalar
• mkomati
• phanitar

We are designing and implementing a Transformers based deep learning model to better capture patterns in demand forecasting tasks(Grocery Sales) and leveraging multi-horizon time series forecasting techniques. The aim is to understand trends/patterns and optimize inventory or to increase sales.

Personalizing AI Avatar Response (Text to Animated Face)

• Devesh
• Nikhil
• Sai Deshith

This project builds an AI-powered avatar system that reacts to user text by generating matching facial expressions. It uses sentiment analysis to detect emotions in text and then creates or edits a face (cartoon-style or user-provided) to reflect that emotion. For example, if the text is happy, your personalized avatar smiles; if it’s sad, the avatar frowns. The system uses modern AI techniques like GANs/Diffusion models to generate and modify facial expressions. The goal is to make avatars feel more emotionally aware and personalized.

Anomaly detection and classification via Object-Centric Pattern Recognition

• Fagun Patel
• Sarayu Sivakumar Dhaya
• Vishali Mani Paramesvari

Project Description This project aims to develop an intelligent traffic monitoring system that analyzes live camera feeds to detect and classify anomalies in traffic patterns. The system will learn normal traffic behavior over time and identify deviations that represent anomalous conditions such as accidents, congestion, vehicle breakdowns, or unusual driving patterns. Unlike traditional traffic monitoring systems that only detect the presence of anomalies, our approach will further classify these anomalies into specific categories (e.g., flow-based anomalies like congestion, behavioral anomalies like erratic driving, infrastructural anomalies like road hazards, and environmental anomalies like weather impacts). This classification capability will enable more targeted responses from traffic management systems. The project leverages recent advances in deep learning, particularly combining convolutional neural networks for object detection with graph neural networks and LSTM architectures to capture spatial-temporal dependencies in traffic data. By processing real-time video feeds from traffic cameras, our system will provide actionable insights that can help reduce congestion, improve emergency response times, and enhance overall traffic management efficiency. Our implementation will focus on balancing detection accuracy with real-time processing requirements, ensuring the system can operate effectively in various environmental conditions and traffic scenarios.

Handwritten Equation Solver: Digit and Symbol Recognition using CNNs

• Vikranth Bandaru
• Thanooj Lingampally
• Narendra Varma Muppala

This project focuses on developing a convolutional neural network based system to recognize handwritten mathematical equations and automatically solve them. Using datasets like MNIST and custom generated symbol sets, we train CNNs for digit and symbol classification. We integrate this with parsing logic to interpret the spatial structure of handwritten expressions and link it to a symbolic computation engine for solving. In evaluation we emphasize on recognition accuracy, models handling of various handwriting styles, and the correctness of computed solutions.

Real-Time Sign Language Recognition Using Deep Learning and Live Video Streams

• Prajesh Gupta Vizzapu
• Jyothsna Devi Goru
• Anirudh Raj Sharma

The project aims to build a real-time Sign Language Recognition system using OpenCV. The goal is to recognize both static hand gestures (such as American Sign Language (ASL) alphabets A-Z) and dynamic gestures (i.e., gestures involving movement across frames or signing complete words like "Hello" or "Thanks") from a live video feed captured via a webcam. This system will process video frames in real-time, detect the user's hand, and classify the gesture using a trained model. For static gestures, a Convolutional Neural Network (CNN) will be used. For dynamic gestures, the project will integrate sequence modeling techniques such as RNNs or Transformer-based architectures to analyze temporal patterns across sequences of frames.

Monologue Prediction Model

• Nate De Lucia
• Chase Allard
• Darren Chiu

Language prediction model trained off of UB lecture transcripts to generate content in a similarly academic style

Text/Audio Based Sign Language Motion Generation

• Rajvi Zala
• Manan Jain
• Rama Krishna Raju Pusapati

We are aiming to generate 3D Sign Language Motions based on the input audio or text.

Wafer Imbalance Normalisation with GANs (WING)

• nancyipo
• arengara
• juliusme

The challenge of data imbalance is critical in high-precision fields such as semiconductor fault detection. Conventional augmentation techniques (rotation, flipping, scaling) fail to capture complex defect semantics, especially for underrepresented patterns like "donut" or "edge-loc". WING leverages GANs to synthesize realistic, class-specific samples for minority defect types, improving model balance and robustness.

YOLO for Satellite Imagery

• Om Atul Salunke
• Swarat Sarkar
• Hamsika RG

YOLO based small object detection with scale adaptations for satellite imagery

Anomaly Detection for Enhanced Public Safety

• Pranjali Sharatkumar Patil
• Sunidhi Patange
• Tejesh Krishna Kadali

Our project focuses on building a smart real-time anomaly detection system for public safety using AI. Unlike older systems that only detect suspicious activities, our model goes a step further—it identifies accidents using YOLOv8, detects the vehicle's license plate through CNNs and OCR, and sends instant alerts to authorities. Inspired by past research, the solution tackles the key gaps: real-time action, accurate vehicle traceability, and automated emergency response. By integrating object detection, behavior analysis, and alert systems, our model offers a more complete and practical approach to enhancing safety through surveillance videos.

(CSE 487: Data Intensive Computing)

MovieML: An Educational Movie Recommendation Tool

• Lexi Zawadzki
• Peter Vaiciulis
• Ryan Cao

A website that simplifies how machine learning recommendations work, helping users make quicker, smarter decisions. With a clean, easy-to-use interface, the app reduces decision fatigue, making it easier to choose content without feeling overwhelmed. As an educational tool, it empowers users to understand how data drives recommendations, using algorithms like K-Means, description embedding, and Random Forest. This improves satisfaction and engagement.

The UI Gen Project

• smotte
• hkorapat

Founders without technical or design expertise often strug- gle to visualize and validate SaaS ideas, leading to costly delays and wasted resources. By simplifying UI generation, this project empowers founders to quickly create and test layouts, turning abstract concepts into tangible prototypes. This not only accelerates idea validation but also reduces reliance on expensive resources, enabling founders to innovate faster and bring their vision to life with confidence.

(CSE 510: Interactive Programming Environments)

Pen-Plotter

• Al
• Tiffany Cai
• Sabrina Johnson
• Nick Brown
• John Abramo

We have repurposed a 3D printer to draw characters on paper using a pen attachment. We wrote code in ZeptoForth, a Forth system designed for ARM Cortex-M microcontrollers, to program a Raspberry Pi Pico 2 (RP2350) to move the printer's motors for precise character drawing.

ACROPALYPSE

• Victoria Dib
• Sam Anderson
• Shawn Zimmermann zimmerm3
• Lizzy Bock
• Gianna Bossoreale

After the world ended, only the grass remained. In ACROPALYPSE, you are a lone survivor tasked with farming sacred grass to appease the Grass God, an ancient entity that demands green life in a dead world. This is a course project from CSE 510: Interactive Programming Environment where we chose to learn about the interactive language Forth.

Project Click

• dylanclo
• dsmuck2

A point and click game written entirely in Squeak Smalltalk.

(CSE 546/676: Reinforcement Learning & Deep Learning )

Inferno Tactics

• Shreyas Bellary Manjunath
• Shaurya Mathur

Inferno Tactics is an end-to-end AI wildfire management system that combines real-time detection, predictive forecasting, and adaptive response. Our hybrid deep models continuously monitor GRIDMET-derived environmental data to pinpoint emerging fire risks, and the moment a blaze ignites, live GPS coordinates are fed directly into a high-fidelity wildfire simulator. There, RL agents—trained on terrain, land cover, and elevation dynamics—rapidly devise and refine optimal attack strategies, allocating resources and tactics in seconds. By fusing proactive risk prediction with on-the-fly, simulation-driven response, Inferno Tactics empowers agencies to outpace unpredictable fire behavior, safeguard communities, and minimize ecological devastation.

(CSE 611: Master's Capstone Project)

Gaitless Feet Height Control Quadrupedal Robot

• Aman Arora

The project aims to train a policy which enables a GO2 quadrupedal operator to control feet height without any designed gaits. If possible we will enable it to climb stairs with a blind policy.

Kaleida Health App 4.0

• M V N S H Praneeth
• Neellohit Somayajula
• Muhammad Waseem Thameem Ansari
• Sri Bhargava Bhamidi
• Mayush Timmapuram

Kaleida Health App 4.0 primarily focuses on improving communication between patients and family, helping patients and their families follow their medical journey, having physician teams provide customized information on diagnosis for patients, and most importantly, decreasing patient anxiety about the hospitalization process.

Appetit Loyalty Program

• Jeremy Lawson
• Rose Vuluvabeeti
• Bhanu Teja Veeramachaneni
• Yue Shao

We developed a loyalty program for Appetit, a food ordering app competing with giants like DoorDash and Uber Eats here in the Buffalo area. We created a user friendly, configurable portal for administrators to create, edit, and delete rewards within a point system. On the customer side, we designed a system that allows customers to earn and redeem points for discounts on orders, encouraging customers to keep using the app and make them feel as if they're saving money.

LED Video Wall

• Alexander Bohosian
• Nicholas Salvemini
• Michael Beards
• Harvey Kwong

LED Video Wall

Dentite

• kbhola3
• anujasun
• skhadtar
• salonika
• sshigaon
• piyushmo

Dentite, a company focused on AI automation in the medical billing sector, we developed a mobile application that streamlines insurance data entry by using OCR (Optical Character Recognition) technology. The app leverages cloud-based OCR models to extract essential fields from insurance cards, such as policy number, provider name, and coverage details, directly from scanned images. This significantly reduces manual effort, saving hours of administrative time and improving accuracy in the billing workflow, ultimately helping healthcare providers focus more on patient care.

Redprint

• Yash Jajoo
• Smit Uplenchwar
• Pratik Fandade
• Vaibhav Sharma
• Zaid Arshad

Redprint Gym App - The interactive gym system

Qu Anytime

• Sruthisri Venkateswaran
• Jackey Wang
• Nawar Khouri
• Darshan Sonawane
• Gurleen Rekhi

Qu Anytime offers growth on demand—for everyone. Inspired by the simple truth that we grow most through experiences and other people, we built a way to access both—anytime, anywhere. Whether you’re on a personal growth journey, eager to explore the world, or craving deep conversation—about the future of our species, your favorite sports team, or a place you’ve always wanted to visit—every topic and call is a chance to connect, learn, and grow with others. Along the way, you’ll develop the Qu Mindset—the ability to learn and grow from every moment—helping you make the most of every experience, on Qu and in real life. Choose a topic. Pick the length of time. Select the number of people. Then connect, learn, and grow—with the world. People you’ve never met. Insights you’ll never forget. Every call offers unlimited opportunities to grow. Qu Anytime is for the curious, the driven, and those who want to realize their potential. Students, creatives, professionals, and everyday explorers—anyone who wants to thrive in the modern world. This isn’t social media. It’s social growth.

Filmic

• Ankit Mukherjee
• Chris Dearing
• Yash Ravindra Sandansing
• Aryaman Mishra
• Praveen Kumar Gangapuram

In collaboration with Filmic, our project focuses on building an integrated web and mobile application that bridges the gap between motion picture archivists and potential funders. The platform enables archivists who possess undigitized motion pictures to create profiles, showcase their collections, and request funding for digital preservation efforts. At the same time, individuals and organizations interested in supporting cultural preservation whether out of personal passion or for tax rebate incentives can browse through projects, learn about historical collections, and contribute directly to the ones that resonate with them

Safety Knights

• Kshithi
• Rakshatha
• Yasaswi
• Pulasthya
• Raghu

The EHS Consulting Network is a web platform designed to streamline how safety professionals hire and manage safety consultants. It addresses key pain points: professionals being overworked, stretched thin, and lacking efficient tools. The platform enables quick consultant hiring through a dynamic job form, automated matching algorithm, and anonymous bid submission system. It also supports consultant vetting and secure messaging. Ultimately, the goal is to save time for safety pros by simplifying the entire hiring and project management workflow from start to finish.

(CSE 676 : Deep Learning)

AI-Powered Traffic Analysis and Management System  Summary

• Harsha Vardhan Meka
• Niharika Vangaveeti

Traffic congestion in urban areas has become a severe problem for cities worldwide, leading to inefficiencies, safety concerns, and environmental issues. This project is designed to create an AI-based traffic analysis system that can detect vehicles in real-time, monitor the movement of vehicles, estimate vehicle speed, detect traffic anomalies (like wrong-way driving), and predict future traffic congestion. By leveraging computer vision and deep learning algorithms, such as YOLOv11 for object detection and LSTM (Long Short-Term Memory) for traffic prediction, this system provides intelligent insights for traffic management.

Personalized Nutrition Advisor: AI-Powered Diet Plans with Integrated Meal Delivery and Expert Consultation

• Praveen Kumar Gangapuram
• Ankit Mukherjee

Discover ‘Personalized Nutrition Advisor,’ an innovative platform harnessing advanced AI technology to create tailored diet plans aligned with WHO health guidelines. Effortlessly input your health profile, dietary goals, and restrictions, and instantly receive customized meal recommendations. Enjoy the convenience of ordering meals directly through the app, with flexible bi-weekly or monthly recurring delivery options to fit your lifestyle. Enhance your dietary journey by easily sharing your generated plans with trainers and healthcare professionals for expert insights and personalized advice. Simplify your path to wellness with seamless meal planning, expert consultation, and reliable AI-driven guidance.

CounselRAG - A Contextually-Aware Legal QA Platform

• Aayush Subramaniam
• Meghna Verma

CounselRAG is a legal question-answering platform that combines Retrieval-Augmented Generation (RAG) with Knowledge Graphs to deliver accurate, context-rich responses to legislative queries. The platform explores how smaller, fine-tuned LLMs can be made more capable by integrating embeddings from user-provided legal documents (persisted in a FAISS vector store) and structured knowledge from a Neo4j graph built from curated Wikipedia articles on U.S. and international law.

Team 24: Vision-Based Intelligent Speed & Tracking Assistant

• anuragra
• mpeshwe
• ishetye

VisionVelocity is a deep learning-powered system that analyzes dashcam video to detect vehicles, estimate their speed, and predict accidents in real time. It uses YOLOv8 for object detection, optical flow with stereo vision for speed and depth estimation, and LSTM with NLP for accident prediction and risk alerts. The system operates purely on vision without relying on GPS or external sensors, offering an intelligent solution for proactive road safety.

Drug Discovery using Deep Learning

• Sushree Nadiminty
• Tejas Dilip Jadhav
• Deekshitha Prabhakar

Traditional drug discovery is costly and time-consuming and can take over a decade and billions of dollars. Modern-day approaches are quicker in this process but are built on molecular graphs as their basic element and do not take into account other rich modes of information and personal variations. This project brings together Graph Neural Networks (GNNs), multimodal learning (molecular structure, biomedical texts, 3D imaging), and personalized medicine to develop a more efficient, explainable, and patient-centered drug discovery pipeline. The goal is to make predictions for drug-target interactions (DTIs), bioactivity, toxicity, and patient-individualized drug response with cost and scalability in drug generation.

Agni Intel

• ppatil23
• swetasah
• anchalda

Wildfires can grow and move rapidly, threatening lives, homes, and natural habitats. Our project builds a tool that looks at satellite images to forecast where a wildfire is likely to spread the very next day. By giving firefighters and emergency managers a one-day-ahead map of potential burn zones, they can plan evacuations, pre-position resources, and issue warnings sooner saving both lives and property.

Handwriting Recognition for Dyslexic Individuals

• Nidhi Pravin Parab
• Sharvari Mayekar
• Grace Evangelene

This project aims to develop a deep learning-based handwriting recognition system specifically tailored for individuals with dyslexia. Traditional OCR models often struggle to interpret the unique writing styles associated with dyslexia, prompting the creation of a custom CNN architecture named AlphabetCNN. The model is trained on a dataset of handwritten English letters that accommodate dyslexic patterns, incorporating targeted preprocessing and data augmentation to enhance robustness. The goal is to support more effective transcription of dyslexic handwriting, promoting greater accessibility in educational and professional environments.

Self-Supervised Learning for Image Representation Learning and Classification of Alzheimer Disease

• Pavithran Gnanasekaran
• Piyush Gulhane
• Rishab Darshan Shylendra

In this project, we are exploring the application of self-supervised learning (SSL) for identifying Alzheimer’s disease from MRI brain scans. We will be making use of a large dataset of unlabeled images i.e, around 86,000 images. We have developed a SimCLR-based self-supervised learning model to learn important image representations through effective image processing techniques and then fine-tuning it with a limited labelled dataset of 450 images for each image class, i.e. Non-Demented, Very Mild Dementia, Mild Dementia, and Moderate Dementia. For comparison, we are also training a classification model using ResNet-50 architecture from scratch. Our goal is to train self-supervised models like BYOL, MoCo, etc. and fine-tune it to get better accuracy. We will be utilizing the pre-trained models, which would perform classification of other MRI scans related to the disease.

(Ethan Blanton 410/510 Section)

Racast

• Nicholas Salvemini
• Mike Beards
• Cole Manzi
• Aiden Ferguson

A peer-2-peer screen-sharing application developed interactively with Racket

(Independent Study / Research)

DriverSight: Dynamic Attention for Distraction Detection

• Krishnakant Sonji
• Shweta Bankar

DriverSight is a real-time system that leverages dynamic attention networks to detect driver distractions by analyzing in-vehicle sensor data. It identifies distraction patterns quickly, enabling timely alerts and promoting safer driving.

Biomarker-Driven Modeling for Early Alzheimer's Trajectory Prediction

• rahuldil

AI and Pharmacy leveraged research project focused on unraveling the early progression of Alzheimer’s Disease by modeling causal relationships between key biomarkers and genetic risk factors like APOE4. Using a nonlinear mixed-effects (NLME) modeling framework inspired by physiologically-based pharmacokinetic (PBPK) models, the project captures individual biomarker trajectories (e.g., ABETA, TAU, PTAU) over time. By leveraging the ADNI dataset and integrating longitudinal data in long format, the model simulates disease progression and identifies early indicators of cognitive decline. This approach enables a deeper understanding of disease onset mechanisms and has potential applications in personalized early diagnosis and treatment planning.

Video RAG

• Karan Ramchandani
• Aditi Sinha
• Vansh Solanki

A Retrieval-Augmented Generation Approach for Video Understanding

Revelio

• bhanucha

AI-Powered Vision with intelligent depth and object navigation system for the visually impaired. We aimed to develop an assistive system that enables visually impaired individuals to navigate their environment safely and independently. By integrating object detection, depth estimation, and speech processing models, the system will provide real-time audio feedback, eliminating the need for external human assistance.

Generative Artificial Intelligence Models for Progression of Speech Disorders in Neurodegenerative Diseases

• Nikita Anil Patil
• Amrut Kulkarni

Many neurodegenerative diseases, including dementia and Alzheimer’s disease, cause progressive worsening of speech. The purpose was to critically evaluate an innovative artificial intelligence strategy for modeling disease mechanisms and the progression of speech disorders.

Systematic Evaluation of Raft using Evaluation-as-a-Service (EaaS)

• sonambar

In this work, I leverage the EaaS framework to perform a comprehensive and reproducible evaluation of the Raft consensus algorithm. By automating workload generation, parameter exploration, and result visualization, this evaluation provides deep insights into Raft’s performance across diverse configurations and workloads. The study not only validates Raft’s efficiency but also highlights nuances that may be overlooked in traditional evaluation setups, reinforcing the benefits of systematic experimentation in distributed systems research.

CRM

• John Abramo

An offline first CRM application for a small business.

Enhancing Satellite Image Resolution Using Super-Resolution Convolutional Neural Networks (SRCNN)

• Chetan Lekkalapudi
• Manasa Chalasani
• Arshad Rathnagiri

In this project, we explore how deep learning—specifically Super-Resolution Convolutional Neural Networks (SRCNNs)—can be used to enhance the quality and resolution of satellite imagery. High-resolution satellite data is often expensive or unavailable due to bandwidth and acquisition limitations. Our goal is to reconstruct high-resolution images from their degraded, low-resolution versions using a data-driven approach. We designed a complete pipeline where high-resolution satellite images are first cropped into 400x400 pixel patches. These patches are augmented through rotation and then degraded via bilinear down-sampling followed by up-sampling to simulate low-resolution data. The model is trained to map these degraded inputs back to their original high-resolution form.

UnCypher

• pmishra6

A multi model navigation platform that assists user that goes beyond just directions, offering insights and recommendations based on user state.

The Nexus App

• Hariharan Sriram

Nexus is a powerful productivity and wellness application that helps users manage their tasks effectively, track their mood, form healthy habits, and unwind with integrated mini-games. Nexus develops a customized path for everyday optimization and emotional well-being using AI insights. It will show how users can easily monitor their emotions, create objectives, remain productive, and take mindfulness breaks using a single, user-friendly platform.

Group 96

• Punit Kumar
• Vaibhav Saran
• Divyesh Patel

Sepsis Treatment Optimization via RL with BCQ+Attention

SAMO:A Lightweight Sharpness-Aware approach for Multi-Task optimization with Joint Global-Local Perturbation

• Gokul Ram Subramani

This project focuses on integrating Sharpness-Aware Minimization into Multi-task learning. We propose a lightweight Sharpness-Aware Multi-task Optimization (SAMO) approach, which combines the benefits of Global Information (G-SAM) and Local Information(L-SAM) to mitigate task conflicts while keeping the computational cost manageable. The effectiveness of our approach was then validated with extensive experiments on different datasets (NYUv2, Cityscapes, QM9, Office-Home).

OneDataShare

• Vamshi Krishna Kyatham

OneDataShare is a web application used for cloning or transferring your data from one cloud service to another cloud service in a very quick and an efficient way.

ActDiffNet: Affective State Recognition using Biological Signals

• Jatin Chhabria

We present ActDiffNet, a novel active learning framework for affective state recognition using multisensor biosignals. The model addresses key challenges in biomedical machine learning, such as data scarcity and class imbalance, by synthesizing hard-to-learn minority patterns using conditional diffusion models. ActDiffNet starts with minimal labeled data and iteratively improves through targeted synthetic signal generation. Our method achieves state-of-the-art accuracy on benchmark datasets, outperforming existing supervised and self-supervised models. This research is accepted for publication at the IEEE CHASE 2025 conference.

Hate Speech Analysis

• saiaakas

This project addresses the challenge of detecting hate speech in online videos using a multi-modal approach. Utilizing the HateMM dataset, the project explores methods for identifying hateful material through the analysis of visual imagery, audio signals, and textual data. Employing advanced machine learning techniques, such as transformer models and audio feature extraction, the project demonstrates how integrating multiple modalities can improve detection accuracy and contribute towards creating safer online environments.

PR: Group 26

• Punit Kumar
• Vaibhav Saran
• Divyesh Patel

Probabilistic Modeling and Transformer-Based Enrichment for Sentiment Analysis on Social Media (Twitter) Data

Toxic In-Game Voice Chat Moderation using Multimodal LLMs

• rsingh57
• nmurali3
• keyanguo
• hongxinh

Multiplayer cooperative and competitive gaming platforms like Roblox, allows more dynamic in-game interactions providing immersive gaming experiences. To increase player communication, they incorporated voice chat features. The chat-features have brought significant safety concerns reported by parents and children on various social media platforms.

Medilink: A Multi-Agent Conversational Pipeline for Evidence-Grounded Symptom Diagnosis

• Bhushan Mahajan

MediLink MultiAgent Bot is an interactive, clinician-inspired chatbot designed to deliver reliable preliminary symptom diagnosis. It combines a MedAlpaca-7B CasualLM (LoRA-tuned) to mimic a physician’s workflow of predicting disease: extracting and standardizing free-text symptoms via UMLS and ICD-10, iteratively querying missing key symptoms through a co-occurrence engine, and grounding its final diagnosis in PubMed literature via retrieval-augmented generation. Dialogue memory is managed with LangChain to maintain context and personalize follow-ups, while confidence thresholds drive an ensemble of discriminative and generative predictions. Upon reaching sufficient certainty, the system issues a chain-of-thought explanation that cites evidence step by step and concludes with tailored precautions. Evaluated on a curated 1000 unique disease corpus, MedLink achieves 95.4 % Top-1 and 99.2 % Top-5 diagnostic accuracy, alongside BLEU 45 % and ROUGE-L 52 % for its generated rationales outperforming traditional classifiers (≤ 52 % accuracy) and single-pass LLMs. By uniting parameter-efficient fine-tuning (LoRA), DeepSpeed optimization, and a dynamic multi-agent loop, MedLink delivers transparent, high-fidelity diagnosis. Future extensions incorporating temporal symptom modeling, multimodal inputs, and expanded rare-disease coverage promise to further enhance its clinical realism and reach.

RD

• rdasari

End to end Audio diarization

(CSE Demo Days)

Self-Supervised and Quantum-Enhanced CNN for Classification & Segmentation of Multispectral Sentinel-2 Satellite Data 

• Hasmitha Krishna Nunna
• Devi Sree Vipperla
• Penugonda Tandava Sai Naga Koushik

This project aims to develop a robust pipeline for land surface classification and segmentation using Sentinel-2 satellite imagery. It leverages self-supervised learning through Momentum Contrast (MoCo) to efficiently extract features from multispectral data without the need for labeled examples. A range of classifiers, including a CNN, Vision Transformer (ViT), and Quantum Convolutional Neural Network (QCNN), are utilized to process the learned features. The project also incorporates a segmentation module using U-Net to mark object-level regions like crops and water bodies. The goal is to create a fully integrated system for land surface monitoring with real-world applications in agriculture, urban planning, and environmental analysis.