Conversational AI agent for Student Services
Conversational AI agent for Student Services
A unified conversational AI platform designed to streamline Georgia Tech student services. This system bridges the gap between static knowledge and dynamic action by integrating a novel Domain-Aware Hierarchical RAG for complex policy reasoning with an Orchestrated Multi-Agent architecture for personalized database querying and schedule management.
1. Domain-Aware Hierarchical Retrieval
Domain-Aware Hierarchical Indexing: Implemented a bottom-up recursive layering strategy using K-means clustering to bridge the gap between high-level policy summaries and specific local constraints.
Cross-Policy Reasoning: Solved the "fragmented context" limitation of standard RAG by synthesizing constraints across disparate documents (e.g., linking employment rules with degree requirements).
Token-Efficient Retrieval: Optimized retrieval performance in low-token regimes, demonstrating superior recall and document recovery compared to baseline leaf-node retrieval methods.
Contextual Synthesis: Enabled the system to understand global context through high-level summaries while retaining access to factual precision in lower layers.
2. Orchestrated Multi-Agent Architecture
Triage-Based Orchestration: Architected a central routing agent that analyzes user intent to dynamically dispatch tasks to specialized Program, Calendar, or Database agents.
Dynamic SQL Integration: Empowered the Database Agent to generate and execute SQL queries in real-time to retrieve personalized student data (GPA, course history) for context-aware advising.
API-Driven Action Execution: Integrated Google Calendar API for direct schedule management (creating/deleting events) and created tools for email drafting and validation.
Stateful Agent Handoffs: Designed a state management system that preserves user context and conversation history during transfers between agents, ensuring a seamless user experience.
Data Analysis of Agriculture’s Environmental Impact
Analyzed and visualized the environmental impacts of agriculture, with focus on biodiversity loss, greenhouse gas emission, and water consumption.
Identified key trends to promote awareness of agricultural sustainability and proposed actionable solutions to address these issues using OECD data.
Developed interactive Tableau dashboards to visualize complex trends, making data-driven insights accessible for sustainability awareness.
Modeling Development of Numerical Concept Sensitivity in CNNs
Demonstrated the developmental alignment between numerosity representation in CNNs and human cognitive development by analyzing distance and ratio effects using VGG-19 pretrained on ImageNet.
Sullivan (using Arduino)
(Educational Device for visually challenged)
Device Drawings
Circuit Diagram
Fast printing speed
Accuracy of printing
Easy to reset previous data displayed
Purpose:
The purpose of the device is to assist visually challenged individuals, particularly children, in accessing and understanding visual information typically presented in educational materials. There are many difficulties for visually challenged children when they study typical visual materials such as maps, diagrams and shapes of animals that we see in textbooks. Even though miniature models can be used as substitutes, there are limitations in covering a variety of educational topics, as it requires either creating or purchasing a suitable product for each lesson. This process consumes a significant amount of time in selecting the appropriate product for even a single class and it also costs a lot. Our team aimed to address these educational limitations caused by such inconveniences. We sought to create a braille device that not only easily incorporates additional visual information for learning but also allows for quick utilization during lessons. Additionally, this device aimed to enable students to review the material on their own.
How does it work?
The Sullivan device facilitates real-time learning during lessons by physically representing visual material in a tactile form that can be perceived through touch. The process begins when an instructor captures an image from a textbook using a smartphone linked to Sullivan and uploads it to the Sullivan app. This image is then added to a database and automatically transformed based on its RGB values. The transformed image is represented by varying pin heights, making it easier to understand through touch and conveniently formatted for output on the device.
Once the image data is named, a user can activate it by voice recognition. Speaking the name of the image prompts the previously transformed data to be transmitted as operational commands to the device's Arduino to operate stepping motors, and solenoids. Consequently, the image is outputted on Sullivan through the pin heights, allowing for a tactile representation of the visual material. This approach enhances the learning experience for those who have difficulty when using visual materials, making education more accessible and inclusive.
Usefulness for real application:
Sullivan is distinct from existing products because user can instantly take and upload photos through a connected smartphone, making it incredibly convenient for use. There's no need to manually convert photos or select device-appropriate images. Moreover, it's highly beneficial in academic settings, especially in subjects like mathematics, where the pace and direction of the lesson often require real-time change in representation of graphs and other visual materials because it is quick in representing the desired visual material. Furthermore, once data is added to the database, existing visual materials can be easily printed using voice recognition, enabling students to use it effortlessly on their own and it makes Sullivan as a tool not only for classroom use.
Heatsink & Laser Actuator Design Project
1. Laser Actuator Designing Project
Restrictions:
Using two motors to realize three motions including vertical & horizontal movement of the laser actuator and cooling fan for heatsink.
Discrepancy between required rpm for the laser actuator movement and rotating motion of fan.
Ideas to overcome restrictions : Realized horizontal & vertical motion simultaneously using one motor and used another motor for rotating fan.
We conceived a vertical drive mechanism using the principle of levers, and for horizontal motion, we envisioned a design where the pivot point of the cannon is fixed to another structure and rotated. To minimize the twisting of the laser line for the vertical drive, we employed a link drive (crank-slide mechanism). For the continuous operation of the fan to be used in the heat dissipation project, we utilized gears for the horizontal drive
Objective:
Design of a drive mechanism to ensure that the laser pointer reaches the given position quickly and accurately during the evaluation
Simulation result
Manufactured result
2. Heatsink Designing Project
Initial Temperature: 𝑻𝒎𝒂𝒙 = 𝟕𝟕. 𝟗𝟑℃
Final Model: LED + Heat Sink + Cooling Fan
𝑻𝒎𝒂𝒙 = 𝟑𝟎. 𝟑𝟓𝟓℃
Manufactured Heat Sink Product
Objective:
The project involves designing and manufacturing a heatsink for LED heat dissipation, through which the process of heat dissipation and weight reduction for a vehicle's LED headlamp is indirectly experienced, and countermeasures are devised.
Process:
In our project, we optimized the heat dissipation of LED through the design of an impinging jet, utilizing direct airflow for efficient cooling. A radial fin arrangement was implemented to concentrate cooling efforts on the high-heat central zone, with fin geometry designed for ease of manufacture and effective air channeling. The heatsink dimensions were determined via ANSYS thermal analysis to balance weight reduction with cooling performance, ensuring significant heat transfer without excessive fin length. Additionally, fin base height was calibrated to enhance the impinging jet effect and facilitate outward airflow, as confirmed by mass conservation calculations.
Result:
The heatsink and fan application on the LED achieved a 45.5% reduction in maximum temperature, lowering it from 78°C to 42.5°C over a 6-minute period. However, there was a variance from the simulated results, with the actual temperature being 12.2°C higher than the predicted 30.355°C, indicating the system's effectiveness while also highlighting potential areas for thermal management optimization.
ServingGo - QR code ordering service
Team
Service Objectives:
Reduce labor costs and self-payment kiosk equipment expenses for restaurant and cafe owners in terms of ordering and payment.
Leverage AI-based recommendations and targeted advertising using accumulated and learned individual customer food preferences and behavior data from completed payments.
Enhancing customer experiences, and aiding restaurant management and revenue growth using data. Technologies include AI-driven recommendation systems and data analytics.
Foster a new culture of ordering and payment in the restaurant industry.
Award:
Hanbat University 13th PSV(Pre-Star-Venture) Entrepreneurship competition Excellence Award and Overseas Training Special Prize (2018)
Service Requirements and Tech Stack Plan
(not officially released)
A. Guest Mobile Application
Framework: Android and iOS applications
Payment Module: I'mport API
B. Owner Web Application
Framework: React.js
Integration with POS and Receipt Machines
C. Server
Framework: Node.js-based real-time network application
Cloud Infrastructure: Amazon Web Services (AWS) (EC2 for server computing resources, S3 for static resources like QR codes and store/menu images, RDS for database storage)
Database: MySQL (for structured data storage related to users and stores), MongoDB (for real-time data storage required for payment data, error logs, and data analysis)
MeCar
Objective:
Since car engines are very heavy and expensive, exceeding the budget we have for manufacturing and the weight capacity of the car frame, we decided to build the car with a motorcycle engine, considering that the performance of a car engine would not be fully realized.
Process:
Our team aimed to compete in a car-building contest for university car club in South Korea, which only teams that can fulfill conditions of sufficient number of members and budget can participate. Despite being recently reactivated after years of inactivity, we utilized minimal resources to attract more members. You welded a frame using leftover pipes from a design studio and purchased a discarded motorcycle engine cheaply from a factory. The budget was primarily allocated to precision-critical components like the gearbox, differential gear, and steering wheel gears. These parts were designed using CAD and custom-made upon request, ensuring their precise fitting and function.
Result:
As shown in the video, we were able to build a vehicle capable of single-person drive. During the university club fair, we gave rides to students and promoted the club, successfully attracting the targeted number of members for the next semester's activities.
Modular Cylindrical Printer Development
(using Arduino)
Purpose:
During my time at school, I often found myself needing to print assignments urgently, but the school's printing facilities were always crowded with other students, making it difficult to handle things quickly. To assist engineering students at my school, my team designed a 3D-printed, portable printer package. This package can be stored in personal lockers and assembled for use when needed. Our goal was to share this design with others to help alleviate the printing urgents we frequently faced.
Key Points:
Designed with a modular structure for storage in limited spaces like lockers, the model can be assembled as needed.
It is intended to be provided as an open-source design allowing anyone to 3D print and use the model themselves.
The product was designed to minimize the elements that users need to purchase. While essential components like Arduino and stepper motors are required, items like pens for printing can be customized as per the user's preference.
Final Product stored in the student locker.
(Realized the goal of the design for limited space storage and modular design)
Independent Study Course Project
Research Abstract:
Multi-layer ceramic capacitors are widely applied in small-sized electronic devices owing to their performance. However, its vibration due to the piezoelectric effect causes the printed circuit board to vibrate and as a result, generates undesired audible noise. By utilizing a simplified finite element model for vibration analysis, the study identified the relationship between mode shape and vibration reduction by changing the positions of multi-layer ceramic capacitors. Finite element analysis confirms that the positions that caused high amplitude are specified by analyzing gradient factors of mode shapes. During the process, representative mode shapes are used to confirm that gradient analyzing results correspond with finite element analysis. The study shows that the best MLCC arrangement for all mode shapes of complex structures can be introduced by analyzing the gradient of the printed circuit board mode shapes.
Research Process
Modal testing to understand experimental model's dynamic characteristics. Adjusting the properties of analysis model to align with experimental model.
(in later process, utilized already developed model to use product of the company)
Modifying the placement of MLCC on the analysis model and performing Harmonic Response analysis.
Research Result
Analyzing method: Selected specific modes
Using harmonic response analysis, it was observed that placing MLCCs on the nodal line reduces vibration, whereas placing them in areas where the mode shape shows large amplitudes results in greater vibration. Based on these results, an analysis of the gradient in the mode shape areas revealed that the largest vibrations occur in regions where the gradient is close to zero. (But both in vertical and horizontal direction)
As a result, concluded that the best MLCC position for the complex structure can be introduced by analyzing the gradient of the printed circuit board mode shapes.
Automation of Claim Matching
(Project as an Intern at Continental Automotive Systems Corporation)
Background problem:
Previous Method: Searching claim detail on client Database manually
Proposed Solution:
Developing algorithm to achieve automation of matching between claim and client database
Three cases organized for comprehensive and effective automation results.
Result:
Estimated annual cost savings of $56,000.
Enhanced efficiency by optimizing ECU quality control and reducing redundant analysis
ECU field claim data monitoring and update from listing system