That’s really impressive. Are they generated in realtime or widgets filled with json?
The UI thing is cool, for sure!
But two Questions came up:
- Why has he not called his girlfriend himself?
- What is the AI talking so long with his girlfriend? May the AI and his girlfriend have a better relationship together, than the girlfriend to him?
The real question should be “why does a developer have a girlfriend?” - because we all know that behind every good developer is a girlfriend that doesn’t exist.
“I need to talk to my girlfriend in Canada… She’s real! She’s real!”
I like the UI here as well. Good example of decision matrix with routing patterns.
Of course with an own binary transport and capabilties of the chat are dynamic - per group or role… and of course with ACL, ABAC, Spatial and temporal and even contract based permission handling…
Widgets are only send to the frontend cache when the user has the right for them.
Locally some answers are so fast that binary transport was the only idea I had to make it even faster.
And only when the system doesn’t know how to handle something it creates a data model extension and workflows and a shell filled with core ui elements which becomes a module that can handle the task with either a decision flow, an agent or preferably deterministically…
Next step is the model training part.. I know I am a little behind on that… Could use more GPU obviously…
How are you doing the reinforcement training?
Yea GPU/compute is king in that area.
I used DeepSeek Harness as a base for my custom harness. I basically rebuilt the internal router and parser, which keeping the look and feel of the app. Deep focus on swarm teams, and am finding coordination the challenging part.
Feel free to contact me to nerd out.
Nah that’s code not AI.. I am trying to replace AI with coding 
I think fresh accounts that have less than a hundret comments on this forum should not be able to use the word Claude. I think you are a bot.
Especially with that username.
@jochenschultz The love your approach of using code to handle things deterministically and bringing AI in only when the system needs to extend beyond what it already knows. I feel like that balance is often overlooked when people talk about AI-powered interfaces.
Exactly something I was discussing other day with the community.
I mean why let the llm use much energy to solve problems when it can just write code to solve something .. which also can be reviewed.. the idea is not really new .. frameworks and cms are build to solve the modularization problem. But I have build a new one with AI having to use it in mind.
I tried with different approaches. Tried wordpress, Symfony, even made a PHP extension to solve issues the language couldn’t help me with back then.
But in the end it is just a data model problem. If you have 10k modules you don’t want to have 100k database tables…