On‑device AI hackathon in Korea: winners and highlights
.png)
Sign up for Developer monthly newsletter
Join thousands of developers around the globe who receive latest news and updates from our monthly curated newsletter.
Sign upWhat happens when passionate developers get their hands on cutting-edge Snapdragon-powered Copilot+PCs? At the Qualcomm Edge AI Developer Hackathon in Korea and Taiwan, teams built smart on-device AI apps that pushed the boundaries of edge computing.
In this article, we’ll explore how developers tackled real-world challenges, what they built, and what they learned along the way.
Following the successful Edge AI hackathons in India and France, a hackathon was also held in Korea in August 2025.
This event involved teams developing Edge AI applications that run on Copilot+PCs powered by the Snapdragon® X series processors and showcasing their results at the hackathon.
The event was organized in collaboration with Microsoft and Korea University and was filled with energy, creativity, and a lot of promising ideas. After a month-long online preliminary period, 11 finalist teams were selected. After completing intermediate training, they returned with upgraded results to demonstrate their skills in the finals.
On the day of the hackathon, the morning kicked off with a demo and mentoring session full of individuality and innovation, and the afternoon was packed with team presentations and Q&A sessions with the judges.
You could feel the effort and passion in every project—students brought their best, and it showed. You could feel the effort and passion in every project—students brought their best, and it showed.
From file management tools to AI-powered email clients, the range of ideas was not only technically impressive but also highly relevant to real-world use cases. The closing happy hour was just as lively, with a surprisingly high turnout— including teams who had already spent the day presenting and even traveled long distances by train to attend the finals.
It was a great chance to connect more casually and get to know the developers behind the projects.
Five teams won awards from Edge AI Developer Hackathon in Korea
1. E.M.Pilot - Team E.M.Pilot
E.M.Pilot is a new AI email client application designed to run entirely on your device, harnessing the power of Qualcomm’s NPU technology for smarter, faster, and more secure email management. Unlike traditional AI email services that rely on cloud processing and come with privacy concerns and subscription fees, E.M.Pilot is open-source and processes everything locally. That means your data stays private, and you get high performance without extra costs.
Key Features
- Smart Classification: Automatically organizes emails into intuitive categories that make more effortless for email management
- Content Summarization: Extracts and summarizes content from email body and attachments automatically (PDF, images, documents)
- Contextual AI Reply Service: Reply your email through native command, email content, and your previous reply to that sender
- Conversational Email Management: Manage your email through conversational commands and intuitive interactions so that you can do use more functions you didn’t know before
- Automatic Calendar Extraction: Automatically identify dates, deadlines, and tasks from emails and save to calendar
- Device Optimization: Specialized Qualcomm CPU, GPU, and NPU implementation for maximum efficiency and performance
Software and Hardware
Frontend
- Tauri: Cross-platform desktop app framework
- React: Modern web framework
- Vite: Fast development server and build tool
- CSS3: Advanced animations and gradient design
Qualcomm AI model
- YOLOv8-Detection: Real-time object detection on images
- Qwen2-7B-Instruct: User command processing and email summarization
- EasyOCR (Detector + Recognizer): Text extraction from images and attachments
- Nomic-Embed-Text: Text intent classification
Backend
- Flask: Backend server
- Transformers: Hugging Face models
- AI Models: Nomic, QWEN LLM, EASY_OCR models
Model implemented
- Email auto-classification (spam/important/sent/to-me/filtering): Nomic-text
- Email summarization: Qwen2.5-7B-Instruct
- To-do (task) extraction: Qwen2.5-7B-Instruct
- Attachment summarization: EasyOCR / Yolo11V / Qwen2.5-7B-Instruct
- AI auto-reply generation: Qwen2.5-7B-Instruct
- Conversational interface (correction/search/calendar integration, etc.): Qwen2.5-7B-Instruct/Nomic-text
Project repo and team members
https://github.com/jinsunghub/copilot_project
https://github.com/rkddlsxo/MailPilot_back.git
Sooun Choi, Intae Kang, Kwanyoung Kim, Jinsung Kim, Sangmin Lee
2. File Fairy- Team FileFairy
File Fairy is an intelligent file assistant designed for your local PC. Instead of relying on external servers, File Fairy uses advanced on-device AI to analyze, organize, and search your documents in real time. That means your files stay private, and you get instant performance—even offline.
Key Features
- Intelligent File Search: Semantic search through document content using AI embeddings
- AI-Powered Renaming: Automatically generate meaningful filenames based on document content
- Real-time File Monitoring: Watch folders and automatically index new/modified files
- Desktop GUI: Clean, intuitive interface built with Svelte and PyWebView
- Multi-format Support: Process PDF, Word, Excel, PowerPoint, text files and more
- Privacy First: Everything runs locally - your files never leave your computer
- Fast Vector Search: Powered by LanceDB for lightning-fast similarity searches
Come for support, stay for the community
Get support from experts, connect with like-minded developers, and access exclusive virtual events.
Software and Hardware
Frontend (SvelteKit_TypeScript)
- Modern component-based UI
- Real-time updates
- Responsive design
Desktop Integration (PyWebView)
- Native window management
- File system access
- System notification
Backend (FastAPI + Python)
- RESTful API endpoints
- File monitoring & indexing
- Vector database (LanceDB)
- AI services (Ollama + ONNX)
Key Components
- Vector Database: LanceDB for efficient similarity search
- Embeddings: Nomic Embed Text model for semantic understanding
- LLM: Qwen3 4B model for filename generation
- File Watching: Real-time monitoring with Python Watchdog
- Text Extraction: Multi-format document parsing
How it works
Intelligent File Name Generation
- Various Document Recognition
- Semantic Analysis / Key Extraction
- Optimal File Name Generation
RAG-based Intelligent File Search
- Real-time Automatic Indexing
- Natural Language Query
- Vector Search
Project repo and team members
GitHub - ben-park2001/file-fairy: your own private local file genie
Byungmin Park, Joon Seo, Sungjin Jeon
3. emerGen - Team UNIDs
emerGen is an Emergency Response Assistant that uses a vector database-powered Large Language Model (LLM) to give anyone instant access to expert disaster response information. Whether you’re facing a fire, collapse, accident, or environmental emergency, emerGen analyzes your voice or text input, retrieves the most relevant guidelines and case studies, and generates step-by-step instructions in real time. The app works entirely at the edge, so you get instant answers without internet or server delays, but it can also leverage online resources for enhanced support.
Key Features
Guideline Generation Based on Redefined Manuals in VectorDB
- When a user provides key information such as their current emergency situation, location, and injury severity, emerGen follows a streamlined process to deliver personalized emergency response guidelines:
- Search a vector database for similar past emergency cases or relevant manuals
- Combines the retrieved data with the user’s input(text/audio). Audio input is also supported by the Whisper-Base-En speech-to-text model.
- Uses a lightweight, finetuned on-device LLM to generate a customized emergency guideline tailored to the situation. By leveraging a vector DB, the system delivers accurate information without requiring additional model training, enabling fast and practical responses.
Q&A service based on Finetuned Llama-3.2-3B-Instruct and Whisper-Base-En
- Users can chat directly with the LLM to ask questions and receive real-time, situation-specific answers related to their emergency. Chat service also supports audio input via the Whisper-Base-En model for speech-to-text transcription.
Keyword-Based Search of Past Cases
- Users can simply enter keywords to search past emergency response cases and manuals stored in the vector database.
- Efficient and Practical On-Device Architecture
- The system operates as a vector DB–driven on-device application, allowing it to function independently without relying on cloud infrastructure.
- Equipped with a small, efficient LLM, it runs smoothly even in environments with limited computing resources.
- Since all information retrieval is handled via the vector DB, no additional model fine-tuning is needed — new data or categories can be added directly to the DB, making the system easy to maintain, cost-effective, and highly practical for real-world use.
Overall Pipeline of emerGen
- Relevant data is pre-organized and embedded into a vector database, categorized by emergency type, task, and context.
- When using the app, the user selects a category, defines a task, and enters a specific question related to the emergency situation.
- If the user provides audio input, it is automatically transcribed into text using the Whisper-Base-En model.
- The system retrieves the most relevant information from the vector database based on the user’s input.
- The on-device LLM receives the retrieved context, the user’s question, and the task type, and generates a customized output tailored to the situation, which is then presented to the user.
Project repo and team members
https://github.com/chaaenni/2025-Qualcomm-edge-ai-streamlit/tree/master
Chaeyeon Jang, Seongmin Lee, Taehwan Kim, Namseok Lee
4. Medly – Team Synaptix
Medly is an innovative solution that leverages the NPU performance of the Snapdragon X Elite to analyze speech in real time and instantly convert complex medical jargon into easy-to-understand everyday language for patients. The project was launched to address the asymmetry of medical information, helping patients gain a deeper understanding of their condition and actively participate in their medical care.
Key Features
- Real-time STT: Instantly converts voice input from the microphone into text
- Terminology Recognition: Accurately identifies and tags medical-related professional terms from the text
- Simplified Explanations: AI analyzes the meaning of recognized professional terms to provide easy-to-understand explanations and summaries
- Comprehensive Summary: Provides a complete summary of the entire voice conversation
- PDF Report Generation: Generates a PDF report that summarizes your diagnosis and allows for printing
- Adjustable Difficulty: Enables setting the explanation difficulty level (Child, Student, Adult) by adjusting the LLM's prompts
- On-Device Processing: All AI computations are processed directly on the device's NPU for maximum speed and security
Software and Hardware
Hardware
- Device: Galaxy Book4 Edge
- Chip: Snapdragon® X Elite X1E-80-100
- OS: Windows 11 Home
- Memory: 16 GB LPDDR5X Memory
- Storage: 512 GB eUFS
- NPU: Qualcomm® Hexagon™ NPU
Software
- Transcription: Live Caption & Tesseract OCR
- Named Entity Recognition: d4data/biomedical-ner-all
- LLM Provider: Qualcomm QNN
- Chat Model: Qwen2.5_7B_Instruct
- Python: 3.12.X (Recommended)
Project repo and team members
https://github.com/laurenlee28/MedLY
Hyunseo Lee, Jooyeob Han, Jaemin Song, Hyeeun Bae, Joon Lim
5. MyStoryPal – Team MyStoryPal
MyStoryPal is an Edge AI-based children's English storybook creation application. It is designed to enable children to learn English naturally and enjoyably through real-time conversations with AI while collaboratively creating English stories. Unlike cloud-based AI services, this app performs all AI computations on the device's NPU, maximizing the advantages of Edge AI.
Key Features
- Conversational Story Co-creation: When children input sentences or ideas, the AI generates appropriate story sentences in continuation. Through this repetitive process, they can gradually complete a story together.
- English Grammar and Vocabulary Feedback: If there are grammatical errors or typos in the English sentences entered by children, the app corrects them. When children ask about word meanings, it provides friendly explanations. Through this feedback structure, children can naturally improve their English skills.
- Creating Personalized Illustrated Storybooks: Every time four story sentences accumulate, the app generates fairy tale-style images that match the content. This creates a more immersive storybook creation experience.
Model implemented
- Text Generation: Llama-v3.2-3B-Instruct
- Image Generation: Stable Diffusion v2.1
- Image-to-Text: CLIP
Project repo and team members
https://github.com/lightsaber1997/story_prototype01/
Gyuyeop Do, Seung U Baek, Seohyun Lee, Yoonjin Choi, Yeonsu Do
Six more teams shortlisted from Edge AI Developer Hackathon in Korea
|
Team |
Description |
Repo |
|
WINE Lab |
AI-powered intelligent secure messenger system |
|
|
Paperclip |
Tone-Sensitive Writing Assistant for Clear and Culturally Appropriate Messaging |
|
|
Seecurity |
AI-based Real-Time Screen Information Protection Solution |
|
|
Famigo AI |
On-device AI Assistant for Family with Voice Interface and Share Memories |
|
|
Team Rocket |
Edge AI Itinerary Planner |
|
|
Jae2 |
AI-based Window Arrangement Optimization System |
Next steps
If you’re interested in exploring these projects further or joining our community, consider the following next steps:
- The projects above are open source, so visit the GitHub repos to use the code in your own apps.
- Take a look through our upcoming events for ways you can see us in person and participate in our hackathons, workshops, conferences and other developer events.
- Join us on the Qualcomm Developer Discord to show off your project and be sure to sign up for the Qualcomm Developer newsletter to be notified of our upcoming developer events across the Unites States, Asia and Europe.
- The Qualcomm AI Hub has more information about using AI models and tools on devices with Qualcomm technology.
