This is a submission for the Gemma 4 Challenge: Build with Gemma 4
What I Built
The Exact Moment It Clicked
Two weeks ago, w...
For further actions, you may consider blocking this person and/or reporting abuse
Splendid execution! Carving out seven low-latency personas from a single base model via Modelfile is exactly how local-first should be done. Controlling the application flow directly through the Blue Hat's output tokens is pure engineering elegance.
No bloated agent frameworks, just minimal loops, queues, and local compounding—this is what a hardcore, AI-native application looks like.
Thank you so much for such a great and detailed feedback. I put so much effort into this, I’m still recovering😄But 100% worth it for feedback like this!
Ah, decision-making loop! Sounds familiar, maybe I should try it 😁
You definitely should. 😄
Now that's the kind of use of AI I respect and support, such cool idea and interesting implementation. Thanks for explaining too, In the beginning of the article I had some questions you all answered perfectly later. Great job brother, really. I'll need some advices later from you, on how to use AI in smart way to stay connected to modern reality but still not lose critical thinking of a good software dev
Thank you! And feel free to reach out anytime, I’m happy to share all the experience I have!🙏
Hey Giorgi, been following your work since the Notion MCP Challenge - NoteRunway was incredible, and this NeuralHats project takes the structured-debate concept to a whole new level.
What caught my eye is the overlap with something I built recently. For the ETHGlobal Open Agents hackathon, I built Deliberate — a Telegram-style crypto group chat where AI agents debate market decisions using structured roles. Different domain, but the same core idea: multiple AI personas with distinct thinking styles arguing through a problem to reach better decisions.
Seeing how you handled the Blue Hat as a controller with the CONTINUE/STOP token, the per-hat temperature tuning, and the structured conversation history to make hats actually engage with each other — that's exactly the kind of architecture problems I ran into with Deliberate.
Would love to connect and potentially collaborate on something in the future. Your work genuinely inspires me and I'd learn a lot working alongside you. Feel free to reach out anytime.
Hey, appreciate that! Tuning those models was definitely the hardes part of the app.
The llm angle here is really well thought out. We've been running something similar with Hermes and found the biggest challenge is actually knowing when NOT to use an agent — sometimes a simple script does the job better.
Great stuff — followed you! 🤝
Thanks a lot! And yes, that’s absolutely the hardest part of the whole thing.
This is a clever use of local LLMs because it does not just ask the model for “better reasoning.” It gives the reasoning process a structure.
The Six Thinking Hats approach is useful here because it forces separation between facts, risks, creativity, benefits, emotions, and process control. That can reduce the usual problem where an LLM blends everything into one polished but shallow answer.
I also like that this works well with local models. For personal brainstorming, decision review, product ideas, or code architecture discussions, keeping the thinking loop local can be useful from both privacy and experimentation angles.
The real test would be consistency. Does each “hat” actually stay in its role, or does the model slowly collapse back into generic advice after a few turns?
That would be interesting to evaluate.
Thank you! And that’s a great question. Teaching those AI models how to think and interact was the hardest part of the application, and for perspective, pretty much every major part of this app was quite challenging.
I spent about 2 days trying to instruct each hat how to communicate and I had to be very specific with the instructions, because otherwise they’d start deviating from both, the topic and the role, big time.
Exactly my kind of idea but I ran mine on my phone
Interesting. I’ll check out yours later. I still haven’t seen other submissions yet.
i can totally relate to the struggle of decision-making loops. having a structured approach like de Bono's Six Thinking Hats is a great way to break free from that. if you're looking to build something quickly, check out Moonshift. you can get a full next.js + postgres + auth app deployed in about 7 minutes, and you own the code on your github. hit me up if you want to give it a try for free.
Any dev here with backend experience
Glad it resonated! The coordination overhead between hats was the surprise — curious if you found certain transitions harder than others when running locally?
One of the hardest part was also making those hats actually interact with each other rather than just throwing out their opinions, you need to be REALLY specific with your prompts. Especially with the blue hat’s prompt, cause that one steers the whole discussion.