Edge Deployments & On-Device AI Hackathon

SafeHome Inventory uses ExecuTorch and Qualcomm's Hexagon NPU to help disaster victims document belongings for insurance claims. The app features real-time object detection (YOLOv8), AI vision-based item identification via draw-to-circle interaction, and on-device LLM-generated descriptions—all running privately with NPU acceleration. Built to demonstrate ExecuTorch's capabilities with Qualcomm QNN backend, it delivers 2x faster inference while enabling accessible emergency preparedness through advanced mobile AI.

Meet Beaker — an app that streamlines experimental protocols for scientists by shortening prep time and providing a spoken interface for step-by-step guidance, note-taking, and progress tracking. Beaker uses on-device Edge AI, so it’s instantly ready to go. Labs don’t need to set up or maintain inference servers, and scientists can use it even in hands-free environments with no internet access. Three easy use cases: upload a protocol document to automatically generate steps; use voice commands for guidance and note-taking during experiments; and track progress effortlessly — cutting logging time by up to sevenfold. Beaker keeps protocol progress stored locally for each experiment, helping scientists stay focused and in their flow state without breaking rhythm.

Have you ever stuttered or forgotten a word mid-sentence? PocketWhisper is an on-device speech assistant that listens locally and fills in missing words in real time - helping non-native speakers, presenters, and anyone who struggles to find the right word at the right time. Built for edge deployment, it runs entirely offline on the Galaxy S25, using ExecuTorch-optimized speech and language models for ultra-low latency inference and full privacy.

Our project (Presentation Buddy) is an on-device presentation coach that gives real-time feedback on how you appear while speaking. It helps users improve their eye contact, posture, framing, and delivery, all without sending any data to the cloud. It's great for teachers, students, introverts, founders, and much more.

Upscale any photo with any zoom captured on the Samsung Galaxy S25 to a DSLR level quality

I got this idea when I wanted all my email analytics in one place but found no private solution. Earlier attempts required sending data to third-party servers, which I didn’t want. With Edge AI, all NLP processing can now run securely on-device. This prototype converts unstructured data like emails and messages into structured insights while keeping everything private. It includes Edge Chat (a developer kit for testing models), Emails Analyzer (extracts information like expenses using a local Llama 3.2 model), and Analytics (the only user-facing app showing categorized spending and trends). Both Edge Chat and Emails Analyzer run silently in the background when the phone is charging. Future data connectors will let apps share data securely for local analysis. All processing happens fully on-device with no cloud or tracking, just private, intelligent insights.

test test test

An automatic motionblur fix for any picture Res

TinyMentor is a Real-time AI Tutor in your Pocket, but for teachers and students in rural or low internet access countries where there are low connectivity and people using low-cost Android devices. Teachers need help with preparing for class but it's hard to ask others. They can send photos in real-time to ask questions about what you don't know in textbook. If students have to walk long time going to school, they can study by using real-time speak transcription while going back and forth from school. to home This AI Tutor in my pocket will decrease the education gap.

Dexa is an on device agent to help you with personal financial information and personal problem, help you keeping your private information private

Recycling is now cheaper, offline, easier, and sustainable. We use Android devices from previous years to run a sorting helper for events, closed spaces, public spaces, and any other location without requiring internet connectivity or expensive hardware.

Does tool calling through Qwen 3 and QNN hardware acceleration loading the model over executorch. The model understands when asked about weather "here" goes to fetch location as a tool call.

Rock gradation analysis app to assist field engineers in geology-related work. Project incomplete (ran out of time), no need to spend time on it.