ZETIC x AWS: Build an On-Device AI App in One Hour
It was Friday evening, in the middle of ICML week. Engineers, researchers, PMs, and founders had every reason to be somewhere else. Instead, they showed up to build on-device AI apps on their own phones.
We called it Build an On-Device AI App in One Hour, hosted with AWS. Minsoo Jo from AWS opened with AWS's perspective on edge AI and how AWS and ZETIC have been building in this space together. Then the room stopped listening and started building.
To make the event itself work across languages, our own team built a real-time Korean-to-English translator, 100% on-device, running on an iPhone. No cloud translation API, no per-minute billing, just Melange running locally, live, for the room.
Cloud and Edge Were Never Actually Competing
The framing he brought to the room cuts against how most people still think about this space. The future of AI is not cloud versus edge. Models get developed and managed in the cloud. Inference runs directly on the device. Operational insights flow back to the cloud where it makes sense.
That's Distributed Intelligence: cloud and edge, each doing the part they're actually good at.
The Same Model, Ten Different Speeds
Here's what most developers don't find out until it's too late: the exact same model can perform completely differently depending on the combination of runtime, quantization, and processor it's running on. Get that combination wrong, and a model that should run in milliseconds crawls instead.
Melange benchmarks every combination across 100+ real devices and automatically routes each model-device pair to whichever configuration actually performs best. Developers don't have to know which combination wins. Melange already tested all of them.
What the Room Built in an Hour
Participants didn't wait for a slide to tell them what to build. One team built a skin condition detector, flagging visible issues directly from a phone camera, nothing leaving the device. Another built an app that turns a photo of your meal into a written review, drafted on-device, for restaurants running review-for-discount programs.
Neither was assigned. Both shipped inside the hour.
The Same Day, We Also Shipped 36 Apps
Alongside the workshop, we launched awesome-on-device-ai-apps, a new open-source repository with the complete source code for mobile AI apps that run entirely on-device. Clone one and run it on a real phone. Or bring your own model to Melange and build a new one in an hour, the same way the room just had.
The first 36 apps in that repository were built by our interns. New apps go up every week.
Why They Actually Showed Up
Most of the non-Korean attendees didn't find this workshop by accident. They came off a week of ICML talks wanting something different, a chance to actually get hands-on and build, not just listen. Some were curious why AWS was talking about edge AI in the first place. So they found this workshop through ICML, and built their travel schedule so the trip ended here.
One attendee put it simply afterward: "Great way to close out ICML week 😊 nice change of pace after a week of talks to actually get hands-on and build something. Thanks to the whole team!"
The deployment problem was never the model. It was always getting from model to device without weeks of engineering work standing in the way. A room full of people spent a Friday night during ICML week proving that gap can close in an hour.
Thanks to AWS and the Edge AI Foundation for making this possible.
Try Melange: melange.zetic.ai
Explore the apps: github.com/zetic-ai/awesome-on-device-ai-apps