The starting point
I wanted a quick way to talk to AI without picking up my phone and getting pulled into other apps. Voice alone was not enough: I also wanted to read responses, mute the speaker, and choose when to use the screen.
The portable prototype uses a Waveshare ESP32-S3-Touch-AMOLED-1.75C board and OpenAI’s realtime voice API. It is a personal experiment in a more focused AI interface.
My role
I direct the product and interaction requirements, test builds on the physical device, and decide which changes to keep. The implementation and code review are AI-assisted. Device testing is where I check whether a change improves the experience.
What the device changed
A caption experiment tied text to audio tags. In an October 27 test, the text arrived 1.8 to 3 seconds late. I rejected that approach in favor of simpler pacing.
A single-word shift in the assistant’s voice mattered more to me than more frequent split-second visual glitches. Frequency alone was not a useful way to rank the problems.
On October 30, an AI review said a recording had no paragraph breaks. I checked the physical device and logs rather than accepting that conclusion. AI review helps, but it does not replace checking the artifact.
Five things to protect
The baseline is an acceptance checklist: start a conversation and get an answer; hear clear audio with matching text; see truthful listening, working, and speaking states; scroll through the conversation; and end a call and start another.
Those are requirements, not a claim that every test has passed. In particular, a screen that says ‘listening’ while a request is working can make progress look like a freeze. Status feedback needs to reflect what the device is actually doing.
Prototype status
In tests recorded on November 1, 2026, muting the speaker kept text responses available, and double-pressing the power button could turn off the screen during a call. The prototype also showed battery and Wi-Fi status and reconnect feedback.
That is a dated test snapshot, not a current reliability claim. Rough edges included a low microphone level, a faint click after audio, slow connection, a transient Wi-Fi warning at startup, and caption timing on short replies. There are no formal usability or battery-life results to report.
More personalized memory and connections to calendars, email, and coding projects are future goals, not capabilities claimed for this handheld prototype.
The next problem it revealed
Testing Luna produced about a hundred videos. Preparing those clips for AI review meant repeatedly compressing them first. That recurring task led to PromptCam.