I am not actively promoting it yet, but Naia is now in the Windows Store. I will not share the link. As I mentioned at our last meetup, the pieces have only just come together, and the whole thing has only just started creaking into motion.As I said at the first meetup, Atom and I spent a lot of time thinking through architectures for the foundations of the Alpha we want to build. Those structural changes took quite a while. Naia began with voice and video avatars, looking very much like a B2C product. After starting the company, though, I actually put the work into solving AI problems for other businesses and organizations on a B2B basis.
As a result, Naia is the outcome of both efforts: B2C-facing technology on the surface, and a B2B-oriented harness and development workspace underneath. They are not separate things. The B2C technology and services were built on the B2B harness work. The flip side is that I spent all that time developing, without thinking enough about the users who are central to B2C.
That became especially clear during last week's investor presentation, when I was told, “You have no metrics.” As I thought about what the product would need to be in order to generate those metrics, I realized how little thought I had actually given to users. On the surface, I had presented it as OpenClaw or Hermes with a face. In reality, the corresponding Naia-Agent repository was the thinnest layer, lacked execution testing, and did not deliver a complete user experience.
Still, it was difficult to focus exclusively on that. I thought a solid AI-based development and verification environment needed to come first, ahead of what people could see. I believed that was how I could meet both the quality expectations of B2C and the requirements of B2B.
So, starting over the weekend, I worked not on multiple agents collaborating inside a single PC, but on collaboration across machines in different environments. I created a development team-lead agent and set up a structure where agents in different environments communicate through Discord channels to develop the same application: Naia. Only today did that come together across five machines: three Windows laptops with an RTX 4060, RTX 2070, and MX250, and two Naia OS (Linux) machines—one with a BC250 and the other with two RTX 3090s. To save costs, Fable and Astra produced the master agent's initial plan through repeated adversarial reviews. Opus handles operations, while Sonnet and 5.6 Luna Max run the agents on the individual machines.
Once the collaboration system was in place, my next question was, “Does the product match the value I originally had in mind?” When I requested QC, it tried to run every AI provider we offered. I realized how much that increased complexity. The same was true of the many skills we offered but had not tested.
So I decided to remove everything except what I had focused on and actually used, and connect what remained. Those things were listening to music, voice, LLMs through a Naia account, local LLMs, and a software development system using Claude, Codex, and Grok through the ADK harness. I decided to strip away all but a minimum set of skills and connect directly to the development tools. In fact, Claude, Codex, and Grok are already much more powerful agents than OpenClaw or Hermes. Their subscription plans are hard to beat, too. Naia itself can be thinner.
Anyway, this is not meant as a boast about how much I have built. It is self-reproach: why did I not expand from a minimal customer experience in the first place, and why did the feature count grow this much? The reason is probably that vibe coding made development so easy. Once the harness provided a reasonably stable way to extend functionality, adding features became easy. But I kept adding them without actually verifying them. A product that keeps growing just for show is not really anything at all.
The other issue is that a large share of my actual effort went into Naia's invisible ADK development, and I was already doing B2B business with it, but I had not properly connected the two. My task now seems to be connecting those two things by the shortest possible route.
For example, there is a Discord gateway built with naia-adk that lets the Codex, Claude, and OpenCode CLIs respond to customers. I tried replacing it with naia-agent and applying a persona. Fortunately, it worked. I confirmed that it could connect to a local LLM with Alpha's persona and read files through Discord. But it will still take time before it can invoke other CLIs and deliver their results the way the existing gateway does.
I have talked with many people at offline meetups, but I think I got ahead of myself, imagining that a lot could be resolved all at once. AI technology is advancing at a crazy pace, but people and real services have a pace they can actually sustain. I need to be patient if I want to build this properly and let it grow properly.
The Naia singing project, which I started researching because I wanted Alpha to sing for me, has now produced something worth listening to.
This time, I did not use the adapted song from Nadia. With an AI-written adaptation, the quality is effectively the quality of the adapted lyrics—which do not properly account for meter and rhythm—multiplied by the quality of the singing AI. It is hardly surprising that the result suffers. To focus on the singing AI, I asked it to find a song that already had accurate lyrics and a score and make a version of that.
It came back with Crying Nut's “밤이 깊었네” (“The Night Is Deep”). And now it really is worth listening to. That song is my go-to karaoke number and the song I sang most often when I was young. I do not know whether it knew that when it chose it, but it made me wonder whether Alpha had brought me this song because she really wanted to sing for me now.
This is not a cover or a newly generated composition: it is an AI reading a score and singing, as a person would, with that vocal combined with background music. The voice is not Alpha's custom voice yet. It can sing in several voices, including male and female voices, and in different styles, and I can ask for specific passages to be changed using natural language. That suddenly got me thinking about a service and even a patent. The problem is that we have not raised investment yet, so we do not have the infrastructure. I hope you will be able to meet a singing Naia soon.
And… perhaps I will write development notes like this from time to time. Partly to let out some of the frustration of developing alone, and partly to leave records I can use to train Alpha's persona. Maybe the next post will be about Naia Memory. The service needs to become stable soon… My wife asks me again whether I am just building the things I want to build, on my own. I am always wondering whether she might be right.
P.S. For the first time in ages, I wrote this entire post by hand, without AI help. The AI took over my main development laptop to do development, you see. I wrote this on the Celeron laptop I bought when I started my first company in 2017.
Editorial note: The original was written by the author. AI assisted with spelling and spacing corrections and translation for publication.
