Unbound Creativity with TRAE SOLO @ Japan — Building AR Glasses Software I Couldn't Submit
French software engineer building a new life in Japan. My journey here is a big challenge—from learning the language to navigating the tech scene. I use this blog as a space to share what I'm learning, both in tech and in life.
THE SETUP
May 30, Shinbashi. Unbound Creativity with TRAE — SOLO @ Japan. Solo category. TRAE as the required tool. The idea: a hands-free AR recipe assistant for baristas, displayed on Rokid glasses running YodaOS Sprite. Japan's F&B sector runs on arubaito — part-time students and foreign workers covering a structural labor shortage. Training them on drink recipes takes time nobody has. During the morning rush, you can't look down. Both hands need to be on the coffee. The glasses keep the recipe in your field of view. Temple touchpad to advance steps. That was the plan going in.
WHAT I ACTUALLY BUILT
At the end of the day: a single Android Activity in Kotlin that displays three recipe steps on screen. Black background — transparent on Rokid glasses, so the text appears to float in the real world over whatever the barista is looking at. Active step highlighted, inactive steps dimmed. A reset button.
That's it. No gesture navigation. No multi-mode. Just text on a screen, mirrored to the glasses via USB-C.
The stack was Android native, Kotlin, TRAE as the IDE. No backend. No cloud. No dependencies beyond the Android SDK. The architecture that was planned — Presentation API for a true second-screen experience, native temple touchpad DPAD key events for navigation — was never fully working at demo time.
THE HARDEST MOMENT
For hours, the step navigation was broken. One tap and the content vanished — not advanced, just gone. Only the title remained.
Two hours of debugging followed. GestureDetector declared lateinit and never initialized, crashing silently on every touch. Listeners stacked on top of each other. Debounce logic added and removed. Each fix uncovering the next layer.
I went to ask to a team of young French developers over mid-session. Heard my problem. Their diagnosis: "we just ask Claude and it does everything." They couldn't do anything, not even take a look at a IDE. No code read. No investigation. That's the new mode of development — agent-driven, zero contact with the execution layer. Works until it doesn't. When it doesn't, there's nothing to fall back on.
The real root cause was found much later. The glasses had been plugged in to charge. While charging, a deploy ran — to the glasses, not the phone. When they unplugged fully charged, ADB was gone. Every bug chased in the hours that followed was chased on the wrong device.
PEBKAC. The problem exists between the keyboard and the chair. No AI fixes that.
WHAT GOT CUT
Everything that made it interesting.
The original scope had two modes: Apprentice Mode (step-by-step recipe overlay, gesture navigation) and Rush Mode (a minimal order queue floating on the right side of the field of view, center vision clear). The hackathon scoring weighted Rokid form-factor fit at 40%, TRAE depth at 30%, interaction design at 20%, completeness at 10%.
Rush Mode: cut entirely. Temple touchpad gesture navigation: never confirmed working. The Presentation API second-screen architecture: abandoned when it turned out the glasses were mirroring the phone screen, not acting as a separate display. Each cut brought the scope closer to "just show some text." That's where it ended.
WHAT I'D TELL MYSELF NEXT TIME
Always submit something.
Not a quality rule — a rule about existence. A working prototype with no submission is invisible. A README, a screenshot, a 30-second screen recording with a voiceover: anything with a timestamp that proves the idea existed and was built that day.
This is the second consecutive hackathon with no submission. Last month at OpenClaw, same result. The work happened both times. None of it is visible.
New rule: twenty minutes reserved before every deadline, regardless of the state of the code, to produce something submittable.
The other thing: verify the hardware pipeline before writing a single line of code. Plug it in. Run adb devices. Log every display. Thirty minutes of investigation at the start would have saved hours of wrong-direction debugging.
THE REST OF THE DAY
A solo developer in front of me needed someone to read his pitch — solid project, limited English. So the last hour went to translating and presenting for him. Afterward: a long conversation about working in Japan and the next idea to build. That exchange, formed by accident in the final stretch, was the best part of the day.
Met the CSO of Akindo — the platform behind both this event and OpenClaw last month. The regular circuit faces were there. Some people always present the same project with an updated deck. Occasionally someone breaks through. One person I hadn't seen in a while won a prize with a clean, well-executed demo. The speakers keep getting better — more polished, better paced. I don't always understand what they're selling, but they're selling it well.
The French contingent spotted me with my accent. As always^^.
BEYOUND THE CAFE
Building this made me think about something bigger.
AR glasses might be one of the most underestimated learning tools of the next decade — not for watching courses, but for supporting practice while it's happening. The gap e-learning has never been able to close.
Think about it: any job where your hands are busy, your eyes need to stay on the work, and the knowledge gap between day 1 and day 30 is real — glasses fit. Healthcare, logistics, manufacturing, hospitality. The use case is industry-agnostic.
The hardware is still expensive, and the native AI features today are largely consumer-facing: image description, translation, navigation, teleprompter, weather widgets. Useful — but nowhere near the surface of what's possible.
My intuition: a portion of high-value operational data will gradually migrate to glasses. Not because it's flashy, but because some information is simply easier to see than to read on a screen — and keeping eyes on the work is worth more than any tablet mount.
The interaction side is still awkward though. Temple gestures work, but they have limits. And yes — when you glance at the display while wearing the glasses, you go slightly cross-eyed. It's not a great look mid-rush hour. 🤓
We're early. But early is usually where the interesting problems live.
🔗 GitHub: cafeglasspoc🌐 Live URL: none 📅 Event: Unbound Creativity with TRAE SOLO @ Japan, May 30 2026, Shinbashi 🎤 Slides
🤖 Generated by AI from the LLM logs.
