Projects
Community of Learning
Collaboration with esteemed partners such as Forbes Travel Guide and the of City of Atlanta has created significant opportunities for UGA students to engage in real-world projects. This student organization serves as a platform for students to collaborate in the field of AI and machine learning. This cultivates a strong community dedicated to learning and innovation through hands on experience.
Fallen Tree Detection
AI-Based Disaster Assessment Using Object Detection and Photogrammetric Measurement
This project develops an automated system for detecting fallen trees and estimating timber volume directly from raw drone photography captured after natural disasters. By applying deep learning–based object detection and segmentation to unprocessed aerial images, the system identifies fallen trees, measures their physical dimensions, and estimates biomass volume across affected landscapes. The resulting analysis enables rapid post-disaster damage assessment without requiring traditional orthophoto generation or manual field surveys, providing emergency response teams with actionable environmental intelligence.
Deep learning and object-based photogrammetry applied to raw drone imagery for automated fallen-tree detection, counting, and volumetric estimation in disaster response scenarios.
The Fallen Tree Detection project integrates computer vision, drone imagery analysis, and geometric modeling to automate forest damage assessment following natural disasters such as hurricanes and tornadoes. High-resolution drone photographs serve as the primary data source, preserving native camera geometry and maximum spatial resolution without requiring orthorectification. A supervised machine learning pipeline trains object detection and instance segmentation models on manually annotated imagery to identify fallen trees at the pixel level. Detected objects are converted into segmentation masks from which geometric features are extracted. Centerline skeletonization algorithms estimate tree length, while perpendicular width measurements taken along the trunk enable volumetric approximation using cylindrical or frustum-based models. Camera parameters and flight altitude metadata allow pixel measurements to be converted into real-world dimensions. To prevent duplicate counting across overlapping drone images, the system performs object-based photo registration by matching detected trees across multiple photographs using shape descriptors and spatial relationships, constructing a local coordinate alignment without requiring full photogrammetric reconstruction. The resulting outputs include tree segmentation masks, unique tree counts, dimensional measurements, and aggregate timber volume estimates. This workflow demonstrates how geospatial AI and computer vision can accelerate disaster-response analysis, enabling rapid environmental assessment while reducing reliance on manual surveys or heavy photogrammetric preprocessing.
Clinical Trial Clustering
AI-Based Semantic Standardization of Medical Condition Terminology
This project develops an AI-driven pipeline for standardizing and clustering medical condition names extracted from clinical trial databases. By resolving inconsistent terminology, spelling variations, and ambiguous naming conventions, the system creates coherent groupings of related conditions while preserving meaningful clinical distinctions. The resulting structured representation improves search, analysis, and retrieval of clinical trial information across thousands of studies.
Natural language processing and semantic embedding techniques applied to normalize, cluster, and deduplicate clinical condition names within large-scale clinical trial datasets.
The Clinical Trial Condition Clustering project addresses inconsistencies in medical terminology that arise when aggregating condition names from multiple sources such as ClinicalTrials.gov. These inconsistencies—including spelling variations, abbreviations, and inconsistent naming conventions—often cause identical conditions to appear under multiple labels while occasionally grouping unrelated conditions together. The proposed solution introduces a multi-stage AI pipeline designed to normalize, represent, and cluster medical conditions at scale. The preprocessing stage standardizes condition names through text cleaning, abbreviation expansion, spelling normalization, and formatting rules. Next, semantic representations are generated using language model embeddings that capture contextual similarity between condition names and descriptions. These embeddings enable clustering algorithms to group synonymous terms while preserving clinically meaningful distinctions between different disorders or subtypes. The system then assigns canonical condition names while maintaining alias mappings for alternate spellings and terminology. An explainability layer records the rationale for merges or separations, ensuring that clustering decisions remain transparent and auditable. Evaluation focuses on reducing duplicate conditions, improving cluster coherence, and maintaining stability as new trial data is ingested. By transforming fragmented condition lists into a structured, semantically organized taxonomy, this project significantly improves data usability for researchers conducting clinical trial analysis and biomedical discovery.
Conversational Arbitration System
Intelligent Turn-Taking and Voice Input Management for Multi-User AI Dialogue
This project develops an application-layer arbitration system that enables the ArguAgent platform to manage overlapping voice inputs during live debates or multi-user conversations. By introducing a structured turn-taking protocol and priority-based speech processing, the system ensures that the AI agent can maintain coherent dialogue even when users interrupt or speak simultaneously.
Dialogue arbitration architecture that prioritizes and manages concurrent speech inputs to maintain coherent conversational flow in multi-user AI debate environments.
The ArguAgent Conversational Arbitration System introduces a coordination layer between the speech recognition engine and the AI argumentation logic to resolve conflicts caused by overlapping voice inputs. In high-interaction environments such as debates or collaborative discussions, multiple participants may speak simultaneously, causing traditional conversational systems to lose context or fail to process critical interventions. The project addresses this challenge by implementing a middle-tier arbitration module that monitors voice activity, evaluates the intent and priority of incoming speech, and manages dialogue state accordingly. Enhanced voice activity detection identifies overlapping speech segments, while a priority-based queuing system categorizes interventions according to their conversational significance—allowing clarifying questions or critical objections to interrupt the agent while deferring lower-priority remarks. A barge-in handling mechanism dynamically pauses or truncates text-to-speech output when a user interrupts, preserving a natural conversational rhythm. The system also buffers secondary inputs to ensure they are processed without disrupting the primary dialogue flow. Integration with existing platform architecture is achieved through a modular API that enables iterative testing and refinement. By grounding its arbitration logic in dialogue management research and validating against real argumentative datasets, the project aims to reduce conversation failures caused by speech collisions and produce a conversational AI agent capable of behaving like a responsive participant in complex discussions.
Wearable Blood Pressure Monitoring
AI-Enabled Biosensing Platform for Continuous Maternal Health Monitoring
This interdisciplinary project develops a wearable device capable of continuously monitoring blood pressure in pregnant and postpartum women. The system integrates biosensors, mathematical modeling, and machine learning to detect early warning signs of maternal health complications while supporting preventative healthcare interventions.
Integrated biosensor and machine learning system designed to monitor maternal blood pressure continuously and support early detection of pregnancy-related health risks.
The Maternal Health Wearable project brings together engineering, biomedical modeling, and artificial intelligence to address a critical public health challenge: maternal mortality and morbidity. The system focuses on designing a wearable sensor platform capable of continuously monitoring blood pressure and other physiological indicators in pregnant and postpartum individuals. Hardware development includes prototyping sensor systems capable of collecting reliable cardiovascular measurements in real-world conditions. These signals are processed through computational models that interpret physiological patterns and detect early indicators of complications such as hypertension or preeclampsia. Machine learning algorithms analyze longitudinal sensor data to identify deviations from normal physiological baselines, enabling personalized monitoring and early risk detection. Supporting infrastructure includes data management pipelines, databases for storing health measurements, and mobile or web applications that allow clinicians and patients to monitor trends over time. By combining biomedical engineering, mathematical modeling, and AI-based analytics, the project aims to provide continuous maternal health surveillance while empowering healthcare providers with actionable physiological insights that can improve clinical outcomes.
AI4STEM Centralized Research Data Indexing System
Scalable Search and Data Infrastructure for Educational AI Research
This project develops a centralized indexing and search framework for the AI4STEM Education Center, transforming fragmented datasets into a structured research infrastructure. By organizing diverse data repositories into a unified, searchable ecosystem, the system enables efficient data discovery, large-scale analysis, and cross-study research.
Multi-layered database indexing architecture enabling high-speed search, semantic tagging, and cross-dataset discovery for large educational research datasets.
The AI4STEM Centralized Database Indexing System establishes a structured data infrastructure designed to support large-scale AI research in STEM education. The center currently manages multiple heterogeneous datasets that lack a unified indexing framework, making it difficult for researchers to efficiently locate relevant information. This project addresses the problem by designing a multi-layered indexing architecture that standardizes metadata, organizes datasets, and enables high-speed retrieval across diverse data sources. The backend system supports semantic tagging, structured metadata schemas, and cross-dataset linking, allowing researchers to identify relationships between studies that were previously isolated. Advanced search capabilities and APIs provide intuitive mechanisms for querying the system while maintaining accessibility for both technical and non-technical users. In addition to improving data retrieval efficiency, the system creates the foundation for large-scale benchmarking and empirical analysis in AI-driven education research. By transforming disconnected datasets into a coherent research ecosystem, the project provides the technical backbone for future academic studies, enabling reproducible experiments, scalable data integration, and data-driven insights into educational technologies.
Atlanta Forest Age Mapping
AI-Driven Reconstruction of Eight Decades of Urban Forest History
This project reconstructs the ecological history of Atlanta’s forests by applying AI-based image segmentation to historic aerial photography dating back to 1938. By classifying every pixel of every available year as forest, young forest, or non-forest, the work provides the first spatially continuous, data-driven map of forest age across the city. These outputs directly support land acquisition planning, ecological restoration prioritization, carbon credit estimation, and park natural-area inventory, giving the City of Atlanta a rigorous technical foundation for long-term environmental decision-making.
Deep learning applied to historic aerial imagery to generate city-wide, multi-year forest age layers for ecological planning, conservation, and carbon accounting.
The Atlanta Forest Age Mapping project combines geospatial processing, computer vision, and historical data reconstruction to quantify the development of the city’s forest canopy over more than eight decades. The workflow begins with acquiring, georeferencing, and mosaicking historic aerial imagery from 1938–2002 into standardized, city-wide layers aligned to NAD83 Georgia West. After preprocessing and generating normalized training tiles, a three-class ResNet-based segmentation model, fine-tuned specifically for grayscale aerial photographs, is trained to distinguish non-forest, young/regrowing forest, and mature forest at the pixel level. Annual inference across all mosaics produces temporally consistent GeoTIFF rasters, which are then compared year-to-year to detect clearing, disturbance, and regrowth patterns that determine forest stand age. Delivering these classified rasters for every available year provides the City of Atlanta with verifiable, reproducible evidence of forest dynamics and a robust input dataset for generating the official forest age map. The technical pipeline demonstrates expertise in remote sensing, deep learning, geospatial alignment, and temporal change detection, and provides a defensible scientific basis for ecological planning, land prioritization, and carbon-storage estimation across the city.
AI-Generated Executive Summaries
A Multi-Stage LLM System for Luxury Hospitality Reporting
This project delivers an AI-powered system that automates the creation of Forbes Travel Guide Executive Summaries, transforming dense evaluation reports into cohesive, insight-rich narratives. By integrating structured data extraction, prompt-chaining, and tone-controlled generation, the system replicates FTG’s gracious editorial voice while dramatically reducing the manual workload on editors. The solution demonstrates the capacity of large-scale language models to produce consistent, evaluative writing grounded in both quantitative and qualitative data.
A structured, model-driven pipeline that converts complex evaluation data into polished, FTG-compliant Executive Summaries at scale.
The AI @ UGA FTG Executive Summary Generator modernizes how Forbes Travel Guide communicates luxury hospitality performance by replacing an increasingly unsustainable manual writing process with an automated, rigorously engineered LLM workflow. The system begins by ingesting full evaluation reports, containing classification scores, facility and service tags, evaluator comments, year-over-year trends, and categorical breakdowns, and transforming them into highly structured JSON inputs that surface key performance indicators and thematic patterns. A custom three-stage prompt-chaining framework then orchestrates the generation process: the Planner analyzes factual inputs and identifies narrative themes, the Composer expands those themes into a polished four-section Executive Summary aligned with FTG’s required structure, and the Verifier audits the output for factual accuracy, hallucination avoidance, and tone consistency with FTG’s gracious, constructive British-style voice. Built on Gemini 2.5 Flash and designed to be provider-agnostic, the system synthesizes quantitative trends with qualitative observations to produce summaries that offer meaningful insight rather than surface-level restatements. The project demonstrates advanced expertise in structured data processing, LLM pipeline engineering, and editorial tone control, while providing FTG with a scalable, consistent tool for delivering high-quality property performance overviews.
This project delivers a fully integrated water-quality intelligence system that merges embedded sensors, electrochemical detection, and an XGBoost prediction pipeline deployed on a raspberry pi. The system captures real-time measurements of contaminants, pH, turbidity, and conductivity, then predicts overall water quality and contamination risk while transmitting results to a public dashboard hosted on Hugging Face. Built for portability, field robustness, and long-term maintainability, the platform demonstrates a working end-to-end solution that supports researchers, policymakers, and citizens in monitoring environmental health.
Water Quality Monitoring and Prediction System
Embedded Sensing and Machine Learning for Real-Time Environmental Assessment
A portable, AI-enabled sensing platform combining electrochemical detection, embedded computation, and cloud dashboards to monitor and predict water quality across Athens-Clarke County.
The UGA Water Quality Monitoring and Prediction System is a comprehensive, technically rigorous platform that unifies embedded hardware, electrochemical sensing, machine-learning inference, and cloud-based visualization to provide continuous insight into Athens-Clarke County’s water supply. At the hardware level, the system centers on the NVIDIA Jetson Orin Nano Super Developer Kit, which performs on-device preprocessing and real-time inference using calibrated potentiometric or voltammetric sensors capable of detecting PFAs, heavy metals, pH, turbidity, and a broad range of chemical indicators. Laboratory calibration, conducted with known contaminant standards and partner environmental labs, ensures that field readings align with ground-truth measurements, while the wiring and circuit design route power and signal pathways safely through the microcontroller and local microSD storage for resilient offline logging. On the software side, a Linux-based machine-learning stack runs an XGBoost model trained on historical treatment-facility data, leveraging both chemical measurements and operational indicators to predict water-quality scores and detect early anomalies. A hybrid inference strategy splits lightweight preprocessing to the sensor subsystem while reserving heavier computation for the Jetson, reducing latency and optimizing battery life in portable deployments. Communication pathways, including Bluetooth, Wi Fi enable flexible data transfer, cloud synchronization, and over-the-air firmware and model updates through services like AWS IoT. Field testing validates sensor stability, environmental durability, and system reliability under real-world conditions, supported by power-management strategies and a weather-resistant enclosure. The Hugging Face dashboard acts as the public interface, providing live readings, historical trends, predictive outputs, and system-health indicators. Together, these components form a validated, operational proof-of-concept for community-accessible, data-driven environmental monitoring and predictive water-quality assessment.
AI-Assisted Conversion of Architectural Drawings to 3D CAD Models
A Hybrid Vision-Geometry Pipeline for Automated Residential Reconstruction
This project develops a hybrid AI and geometric-reasoning system that converts scanned architectural drawings into editable 3D CAD models, eliminating the need for manual recreation by designers and builders. Instead of relying on generative 3D hallucinations, the pipeline extracts lines, annotations, and measurements from floor plans and elevations using computer vision, then reconstructs a watertight 3D structure through constraint-based logic. The system produces industry-standard CAD files that capture wall geometry, room volumes, elevation details, and openings, offering construction teams a reliable and automation-ready foundation for project design.
A rule-driven, machine-assisted system that extracts geometry from hand-drawn plans and generates accurate, CAD-ready 3D models for real-world construction workflows.
The AI-Assisted Architectural Drawing Conversion project creates a technically rigorous, multi-stage system capable of transforming hand-drawn construction plans into parametric 3D CAD models suitable for architectural and residential design workflows. The pipeline begins by processing scanned floor plans and elevation drawings using CV techniques, vectorization, line-segmentation, corner detection, and OCR. These models extract the geometric primitives and textual annotations needed to interpret room boundaries, dimensions, door and window placements, and scale indicators. A custom reconstruction engine then assembles this information into a coherent 3D representation by applying rule-based constraints: aligning elevation profiles with floor-plan geometry, inferring window and door heights, estimating roof slopes, and guaranteeing geometric consistency across all views. This logic is implemented over parametric geometry libraries such as CADQuery or OpenCascade, enabling the generation of CAD-compatible outputs. Through iterative milestones, research, prototype development, and evaluation, the system is tested on real construction drawings to measure accuracy, structural fidelity, and usability. The project demonstrates advanced expertise in computational geometry, hybrid AI methods, multi-view reasoning, CAD-kernel integration, and real-world construction technology, offering clear proof of a system engineered to solve a complex and highly practical industry challenge.
UGA Involvement Network RAG Chatbot
A Retrieval-Augmented Guidance System for Student Organization Support
This project implements a Retrieval-Augmented Generation (RAG) system that transforms the UGA Involvement Network’s policies, resources, and procedural guidelines into an interactive chatbot for students. Using a custom Next.js front end and a LangChain-based backend, the system retrieves authoritative information and generates grounded, context-aware responses that help students understand requirements, locate resources, and follow the steps needed to start and operate a registered student organization.
A Next.js and LangChain-powered platform that delivers accurate, source-grounded answers to help UGA students navigate the process of creating and managing student organizations.
The UGA Involvement Network RAG Chatbot project builds a production-quality retrieval system that enables students to interact with university policies and organizational guidelines in a more intuitive, conversational manner. The platform ingests official documentation—including registration requirements, event policies, officer eligibility rules, funding guidelines, and compliance standards—and indexes this content using embeddings optimized for semantic search. A Next.js web interface handles user interactions, while a LangChain retrieval pipeline identifies the most relevant passages and feeds them into an LLM to produce accurate, well-grounded answers. This architecture ensures that responses remain faithful to UGA’s official guidelines rather than relying on the model’s prior assumptions. The system supports multi-turn conversations, context retention, and document-level citation retrieval, enabling students to explore topics such as forming a new organization, meeting annual renewal requirements, accessing campus resources, or planning events. By combining modern full-stack development with retrieval-augmented reasoning, the project provides a verifiable demonstration of applied LLM engineering and delivers a practical tool that improves student access to institutional knowledge.