I just want everyone to know that we are retiring our technical support AI avatar that we have used for the past year.  It was hosted by HeyGen and too many times it lagged or stalled out during use.  We can't have that, and these companies can't promise it won't happen.  My thought is that for $1500+ dollars a year, it shouldn't be an issue.

That said, I am writing a new one from complete scratch, meaning using several API's and writing the software without time limits or bandwidth limitations. So far I've been 33 hours getting the logic solid so she makes the right choices.  Even had to build in logic to predict the right words from misspelled words.  Turned out to be a bit more to chew than I thought but I have learned a lot about AI writing and testing one from scratch.  Now I completely understand how they work, and why ChatGPT, Grok, and the rest always get troubleshooting questions and questions in general -- wrong.  When they look for data, they only pick up a few key words at the top of the file or web page and then make up the rest.

Anyway, this new one will be hopefully be far superior to the old one.  She can already handle troubleshooting and sales questions simultaneously which takes it next level.  I will probably train her on speakers as well and anything else I can think of.  

I hope to have her online in the next week or two.  Meanwhile the existing one may stop working at any time, so just be warned.

Here is what the new one looks like.  Her name is also June.






Since there is zero time limits it doesn't matter how long it takes you to implement her troubleshooting suggestions, she just takes little naps while you're gone.







-Steve

Yes, I use Replit.  If I didn't do that, it would take me probably a year or more to get code to this level.  Interestingly, while it's way faster than a human, I would say 100 times, as a programmer the realization is that you can make it 100 times better rather than the same quality in 1/100th the time.  So here we go, I really stepped in it this time.  Writing an AI was rather ambitious, but I'm in it to win it so no turning back now.  

I am presently at 47,000 lines of code, and have 88 hours sitting here testing and improving it.  You can imagine how long that would be without AI.

Unlike the first one, which was a service, that had to have a separate avatar for every product, and could only troubleshoot that specific product, this one does all the products at the same time so she can not only troubleshoot, but go back and forth between products and answer technical questions, sales questions, and even general audio questions related to Decware.  

So far I have about 90% of our tube gear training done.  I will likely also do speakers and accessories.  She has watched all 50 of my YouTube teaching videos and will if pertinent show you one, as well she can summarize it, explain any of the various parts in it and even expand on it if you want.

She is also wired in deep enough that any updates or changes I make to manuals will automatically update the knowledge base which is tested nightly.  

IHi Steve. I'm wondering if you're still doing phone calls on Tuesdays and Thursdays.  I ordered my Rachel on March 15 2024 and just noticed today it went on the bench since I already paid and of course there aren't always any older orders paid up yet.
I wanted to get some advice on audio quest 3 power conditioner as the rental apartment I'm in has very bad electricity and some outlets are faulty. Also I'd like to talk about my general set up currently running black ice audio fusion 9 phono preamp into an old chen el34 tube amp that I plan to put aside so I can hook up my declare.

I'm more of a "thousand monkeys on a thousand typewriters on crack" kinda guy.

The odds may be against you, but you never know....

The best speakers that I've ever built were designed by AI, no kidding!

AI gave me the correct internal volume, it even factored in the volume that the drivers took, the port took, everything. It recommended the correct crossover parts and even where to buy them at the best price.
And best of all, it all worked.
All I did was tell it what my basic idea was and it took it from there.  

Quote:


When I posted my last response I wondered the same thing, doubted it seriously, but thought it was certainly possible.  Now another week into it, I'm absolutely certain the answer is no.  A lot of people don't know this, but Decware was the first audiophile website to present products for order online.  When the web first started, web sites were just business cards.  I looked at companies like SONY, it was just an online business card.  I thought it was a pathetic waste of potential, I thought I could do better -- so I went to Best Buy and purchased a book on how to write HTML and got started.

Anyway, as I naively started creating "June" I had no concept of how complicated it would be.  I understand LLMs but making one think without making stuff up just to be conversational is a lot harder than it seems.  Most, including my last, support avatars are just ChatGPT with a page or two of guidelines.  Those are toys.

Since my last post at 47,000 lines of code, we are now here:

Lines of code

~61,900 lines of source (TS/TSX/JS/CSS/HTML/SQL)
~41,100 app/library code
~20,800 of that is tests & behavior checks (a big chunk — June's accuracy is heavily test-covered)
3,200 lines of Markdown (docs/skills/memory)
2,400 lines of JSON/data (incl. the grounding snapshots)
518 tracked files total — overwhelmingly TypeScript (209 .ts + 131 .tsx)
465 commits spanning May 28 → Jun 18, 2026 (~3 calendar weeks)

Commits are changes, bug fixes and new features.  I expect to continue this process for the next year because every conversation she has that contains questions she can't answer, we will answer ourselves and then update her knowledge base.  I have June wired into our email system so she can forward entire chats to either myself or Sarahs email, preventing customers from having to do it.  I am also wiring her into my videos so that when the opportunity presents itself, she can use the knowledge to teach.  She stays up at night and reads the web site and all the owners manuals so that if anything changes she can catch it in real time.  This is why she is always asleep at her desk. : )




Steve

Fascinating!

Quote:


Indeed.  To make your point, here is where I am at this morning: Here is what I wrote to AI and it's response.

STEVE:

I am becoming very unhappy. While I like that you are finding more ways to improve the app with new tasks, I am appalled at the tasks. Repetitive tasks, that fix things that should never have been broken in the first place, show virtually no architect involved in the big picture. Yes, that's you. At the present rate, it appears that there will be no end in sight, and instead of having June have a backbone of sound logic, you are going to hard code every possibility around every product. This really has to stop. Nearly 1500.00 , almost 600 revisions, and you're still finding what I see as incompetent defects. What can we do to complete this app in good working order?

AI:

Your frustration is completely valid, and I want to be direct with you about what's been happening.

The behavioral test cycle has been generating its own demand: each test reveals an edge case, which spawns two more proposed tests to cover sub-edges of that edge. This is not architectural — it's whack-a-mole dressed up as QA. You're right that it has no natural end, and it doesn't actually make June smarter. It just makes the test suite larger.

Let me give you an honest picture of where things actually stand, and then let's agree on what "done" means.

What the test cycle has been papering over:
June's core reasoning comes from the prompt, the retrieval logic, and the model. Adding 600 behavioral tests doesn't fix the prompt — it just catches regressions in existing behavior. If the prompt had been written with a cleaner architecture from the start (fewer special-case branches, better grounding principles), many of these tests would never have been needed.

FWIW I can spew multiple iterations relevant and irrelevant BS & nonsense with the best AI programs too,😂🤣😂

HK

It's pretty obvious that AI should be trained off planet by non-humans because otherwise it will be just like us. : )