source: pytorch blog: building the future of on-device ai at the executorch hackathon
level: technical
a recent hackathon in san francisco brought together over 100 developers to build on-device ai applications using executorch on samsung galaxy s25 ultra phones. the two-day event focused on creating apps that run locally, without cloud dependence. teams worked with pytorch models and received support from qualcomm and meta experts. the goal was to show how ai can work directly on mobile hardware, handling tasks like vision and speech with low latency and high privacy.
the winning projects highlighted practical uses of edge ai. first place went to safescreen ai, an app that detects and blurs harmful visual content in real time on the device. second place, sixthsense, used a phone camera and haptic belt to help blind users navigate by feeling obstacle directions. third place, toddle ai, analyzed toddler walking patterns from video while keeping sensitive data private. all three relied on local processing to deliver fast, private, and reliable experiences that wouldn't work as well with cloud-based ai.
the event showed that developers are ready to move ai beyond the cloud. executorch, part of the pytorch edge ecosystem, lets models run efficiently on cpus, npus, and other phone hardware. participants dealt with real constraints like battery life and memory, proving that on-device ai is practical for products needing quick responses, offline access, or data privacy. the hackathon demonstrated a growing shift toward hybrid ai, where local and cloud processing work together to meet user needs.
why it matters: on-device ai reduces latency, protects privacy, and works offline, making it essential for real-world apps in healthcare, accessibility, and safety.
source: pytorch blog: building the future of on-device ai at the executorch hackathon