Julian Neagu

the idle time idea is real, but dev tools are super sensitive to anything that feels like ads in the flow.

this is the first time i’ve seen someone put agent boundaries and audit trails before models. that order matters. most teams still jump straight to rag + prompt and wonder why clinicians don’t trust it

this is the part most teams miss dashboards aren’t for showing data, they’re for making fast calls under stress. if it doesn’t help answer “is this bad or not” in a few seconds, it’s just decoration.

this is one of those “simple idea, hard execution” tools. turning menu bar space into a live schedule strip actually makes sense for focus work.

this request vs response split is the part people usually miss. once you treat request as “what model sees” and response as “what system allows”, debugging gets way easier in real flows.

Most people still imagine ChatGPT as actually thinking. It’s really just predicting the next token based on patterns.

The HTML over the wire part stood out to me. Fewer moving pieces usually means fewer weird bugs, especially when you're building and maintaining everything yourself.

I like that you kept this simple. A lot of people think the model is searching the web every time. That one idea alone clears up a lot of confusion.

The deterministic layer idea makes sense. LLMs for intent, not for final correctness.

AI workflows feel simple until they sit between payments, users, and internal tools. That’s usually when control starts to matter more than convenience.