From classroom to symposium: Hansung university students create AI-powered hardware solutions

Users can tap a button and issue a voice command to request information, which enables safer and more independent walking. For example, at the voice prompt “Let me know if there is a cafe nearby,” the app uses the LLM to extract the keyword “cafe.” It uses on-screen text OCR recognition to find nearby matches and returns voice guidance on distance and direction.

Expected effects

The team expects that AMASVI will substantially support independent and safe walking for people with visual impairments, enhancing autonomy and improving quality of life. Future integrations with GPS and indoor location recognition will add indoor navigation and the ability to request assistance in an emergency.

Project repo and team

https://github.com/HSU-Capstone-Design/AMASVI

Yoo Geon-woo, Lee Sang-yoon, Baek Seung-heon

2. BrainTrace – Team 무무

BrainTrace, a project built for Windows on Snapdragon, uses an on-device LLM and knowledge graphs to analyze PDFs, notes and voice data locally. This AI assistant is like a second brain that not only automates the process of organizing the material but also executes locally for greater privacy and security.

The user asks a question in natural language. The on-device LLM internally converts the question into a Cypher query, retrieves relevant information from the Neo4j database, and provides a source-based, contextual response. BrainTrace goes beyond simple document analysis tools by applying multi-level inference and connection data to understand the questioner's intent.

Main features

• Runs locally, without the cloud
• Extracts information automatically when user uploads PDF, text or voice
• Creates concepts (nodes) and relationships (edges) using natural language processing and stores them in a Neo4j graph
• Provides intuitive graph exploration based on Cytoscape.js
• Simplifies context tracking with node click/highlighting
• Uses retrieval-augmented generation (RAG)-based, contextual Q&A
• Transforms user questions into Cypher queries via on-device LLM, then extracts responses from Graph DB and provides sources

Technology stack

• Front end – React, Cytoscape.js, Vite
• Back end – FastAPI
• Database – Neo4j (Embedded), SQLite
• Voice recognition – OpenAI Whisper
• AI model – Llama, KoE5
• RAG configuration – Natural language → Cypher → Graph response
• Other tools – Electron, VS Code, Git

Expected effects

The team sees the value of BrainTrace as an AI learning assistant in education; for example, significantly improving the quality of learning from lecture materials. In the enterprise, BrainTrace can help with knowledge management by automatically improving search efficiency in meeting minutes and documents while preserving privacy and security.

Project repo and team members

https://github.com/HS-MUMU/BrainTrace_OnDeviceAi

Kim Dong-hyuk, An Ye-chan, Yoo Jeong-gyun, Jeong Woong

Demonstration video

3. Wakey – Team Wakey

Wakey is a smart album service that creates travel journals using metadata from photos stored on a smartphone. It uses on-device AI to automatically tag photos, then allows users to easily find and manage their memories through natural language queries and timeline creation.

The app is built on the Samsung Galaxy S25 (Snapdragon® 8 Gen 3) and Qualcomm AI Hub, with all AI processing performed on the device for fast, private, secure operation.

Main features

• SmartTag takes advantage of on-device AI to provide tags quickly, without using the internet. It applies YOLOv8 to images for object detection, hashtag generation and cropping of detected objects, then MobileNet V3 to perform classification. Examples: #Motorcycle, #PhotoSpot, #Shop, #Umbrella, #Soup
• SmartSearch provides fast, accurate search and retrieval by matching images with natural-language embeddings. It uses the OpenAI CLIP model to convert images and text into 512-dimension vectors for matching, then enables image search through natural language queries. Examples: “a beautiful sunset at the beach,” “yummy food with beer,” “pork on a grill.”
• SmartStory automatically creates a timeline based on shooting location, time stamp and tags, then generates stories with selected photos and inferred route maps. SmartStory can use the ESRGAN model to upscale low-quality images.
• SmartAlbum uses location metadata to automatically organize albums by domestic, overseas and regional categories.

Project repo and team members

https://github.com/HSU-Wakey

Yang Jun-young , Nam Yoon-chang, Choi Eun-seo, Kang Min-seo

Six teams presented at Qualcomm University Relations (UR) Platforms Symposium

Six of the projects were selected for presentation at the Qualcomm UR Platforms Symposium 2025.

1. AMASVI (A Multimodal Assistance System for the Visually Impaired) – Team 길동무

Described above, AMASVI received an award from HSU and was selected to present at the symposium.

2. Sporty Up – Team Sporty Up

Sporty Up is a mobile system providing real-time posture assessment and feedback.

Developed using Qualcomm AI Hub, it offers on-device AI for instant posture correction during physical activities, along with server-based analysis for precise results. The system also features a community platform where users can share experiences and feedback. This hybrid approach ensures efficient performance on mobile devices, offering a convenient, real-time solution for improving posture and supporting indoor sports activities.

Specifically focused on bowling, Sporty Up implements real-time analysis and feedback of pitching motions into a mobile application. Optimized for mobile devices through lightweighting, Sporty Up improves inference speed and memory usage for real-time operation in real-world environments.

Sporty Up uses the YOLOv11 object detection model and pose estimation model to recognize the user's posture. It uses the ChatGPT-4 API to interpret posture and provide feedback in natural language. To accommodate the resource constraints of the mobile environment, the YOLOv11 model was quantized (W8A8) into 8-bit fixed-point and INT8, respectively, using Qualcomm AI Hub. The result was converted to TensorFlow Lite format for mobile optimization on a Samsung Galaxy S25 powered by the Snapdragon® X Elite.

Project repo and team members

https://github.com/HSU-capston

Son Ju-wan, Lim Seung-taek, Jang Woo-jin, Hwang Jun-hyun

3. StoreShield – Team StoreShield

StoreShield analyzes video with real-time AI to provide theft prevention and inventory management for unmanned stores.

The project aims to use real-time customer tracking and ID-based behavioral analytics to reduce the risk of theft. Its automated inventory management system is designed to prevent out-of-stock conditions and make ordering more efficient.

System architecture 

Main features

• Security – StoreShield automatically assigns a unique ID when a customer enters the store, allowing real-time tracking of in-store movement. It confirms payment status at the kiosk and sends a theft alert if an unpaid customer ID disappears from the screen.
• Inventory monitoring and ordering – Inventory is automatically updated at the payment kiosk, with notifications when low-stock thresholds are met. The system recommends an optimal order quantity based on inventory data.
• Store management – Managers enjoy streaming of live footage from the sales floor, with notifications and remote response. They can analyze daily/weekly/monthly sales data in real time and study customer behavior by time of day, dwell time and product placement.

Project repo and team members

https://github.com/Store-Shield

Yoon Hyun-do, Lee Seon-bin, Jeon Sang-woo, Hong Hye-chang, Choi Eun-seo 

4. MusicGen for Android – Team Android MusicGen

MusicGen extracts emotional information from images and generates music based on it. The app uses Transformers.js to convert text to music in a browser.

MusicGen for Android is an AI-based creative tool that automatically generates music by extracting emotional information from images. Users simply upload a single photo, and the AI interprets the mood and sentiment of the image to generate an emotional description that includes genre, atmosphere, and BPM. Based on this description, the MusicGen model then automatically composes creative music.

Project repo and team members

https://github.com/ChoIntelligence/MusicGen-mini

Yoon Hyung-sik, Cho Jae-hyun

5. Deep Voice Phishing Detection – Team 김미영팀장

Deep Voice Phishing Detection is a mobile security application developed to counter voice phishing crimes using deep voice technology. It can detect deepfake audio and phishing conversations in real time during phone calls. The system fine-tunes the Wav2vec2 model with Korean voice data to determine whether the audio is forged. It also analyzes conversations by comparing them to actual phishing cases using a Whisper STT and EXAONE-based RAG architecture. The application automatically sends recorded call audio to a server in real time, where it is recovered and analyzed, and then alerts the user if a threat is detected.

Project repo and team members

https://github.com/leader-kimmiyoung/deepvoicephishing

Kim Min-sang, Kim Min-seon, Jang Hyun-gyeom

6.  Hazard Sound Detection Application for the Hearing Impaired and Elderly with Hearing Loss – Team 4조

A hazard sound detection application for the hearing impaired and elderly with hearing loss detects dangerous sounds in real time and provides visual and tactile alerts to users with hearing impairments or age-related hearing loss. It integrates Google’s pre-trained YAMNet model, fine-tuned with the Urban Sound8K dataset, into an Android application. By analyzing microphone input in real time, the application identifies hazardous sounds and triggers screen alerts and vibration notification when such sounds are detected.

Team members

Lee Yoo-jun, Ryu Ji-hoo

12 more teams built their Capstone Design projects around technology from Qualcomm

The following twelve teams also participated in the integration of products from Qualcomm Technologies into the 2025 Capstone Design course, through close collaboration with Qualcomm Korea:

The voice of the participants

Park Young-bin, Undergraduate student, HSU

“We put a lot of thought into dynamically linking UI behaviors and managing data flow and state between activities to create a seamless user experience. Furthermore, applying AI models to real-world apps, we experienced the importance of mobile optimization within realistic constraints like model lightweighting and resource management. Ultimately, we felt a great sense of accomplishment in enhancing the app's completeness.”

 Yoo Doo-yeon, Undergraduate student, HSU

“During the AI testing and LLM module integration process, I was able to directly address issues specific to on-device environments, such as model loading, inference speed, and memory issues, which proved invaluable. Initially unfamiliar, I gradually mastered the JNI and TFLite model structures, gaining experience connecting them in a form suitable for actual services. This project provided me with a hands-on experience of the entire process of integrating technology into a product, which was incredibly valuable.”

Seo Yeon-soo, Undergraduate student, HSU

“This project allowed me to experience the entire process, from UI/UX planning to actual implementation. Resolving conflicts that arose during the process of implementing the design into code greatly emphasized the importance of communication during collaborative processes.”

Seong Si-woo, Undergraduate student, HSU

“Focusing on the design and structure of the local database (SQLite), I was able to devise a structure that reliably processes actual user data. Rather than simply implementing functionalities, I focused on creating a maintainable structure. Through trial and error, I also developed debugging and problem-solving skills through testing integration with AI features.”

Namyun Kim, Dean of the College of IT Engineering, HSU

“With Qualcomm’s technical support and mentoring, students gained hands-on experience with On-Device AI, optimizing models such as LLM, Whisper, and MobileNet for mobile devices. This project enhanced practical skills in model compression, data processing, and system integration, demonstrating the potential of Qualcomm’s technology to support both talent development and industry innovation.”

Heeseok Oh, Assistant Professor, Department of Applied AI, HSU

“Hansung University’s Department of Applied AI aims to cultivate interdisciplinary talent with expertise in AI algorithms and embedded intelligent IoT. The Capstone Design project using Qualcomm chipsets and SDKs aligned well with our educational goals and yielded excellent results through students’ active participation. By engaging in real-world applications, students advanced their professional capabilities. We sincerely thank Qualcomm engineers for their dedicated mentoring and contribution to the educational impact of this initiative.”

Danny Jeong, Senior Director, Qualcomm

“We at Qualcomm are sincerely grateful for the chance to collaborate with Hansung University on this Capstone Design project. Witnessing the innovative solutions developed by the students has been truly exciting. Their impressive achievements clearly showcase their dedication and creativity. “

Shaun Kim, Senior Engineer, Qualcomm

“It was a great pleasure, as an AI Software Customer Engineer at Qualcomm, to work directly with the students in implementing on-device AI services using Qualcomm’s excellent AI development environment. I’m confident that this hands-on experience will be highly valuable for their education and will contribute meaningfully to future development of GenAI services based on Qualcomm chipsets in real-world applications.”

Your turn

Besides resulting in amazingly useful applications, the Capstone Design program and Qualcomm UR Platforms Symposium provided students with recognition that could aid them in their job search.

Here are some next steps you can take:

• The projects above are open source, so visit the GitHub repos to use the code in your own apps.
• Join us on the Qualcomm Developer Discord to show off your project.
• The Qualcomm AI Hub has more information about using AI models and tools on devices with Qualcomm technology.
• Our University Relations Program invests in more than 100 university research projects annually, conducted by faculty and students at the undergraduate, graduate and post-graduate levels. Contact us about collaborating with your university.