Appreciate the push on this ,and it’s a fair question. However , I would like provide bit detail about my setup, Which entirely Application layer based.
It spans multiple layers.
In this design, the handling is primarily done at the application layer rather than relying on GPU or infra-level control.
The approach is:
User input is intercepted via a middleware (Cloudflare Worker)
It’s transformed / abstracted before inference (PII minimization)
The model processes only that transformed input
The response is scrubbed again on the way back
We don’t persist raw user text , only derived signals
So the goal here isn’t to claim zero exposure at the infra level, but to:
minimize identifiable data before it leaves the application boundary
and avoid retaining any raw input in our system
It’s more of a transient processing + non-retention model than infra isolation.
Still early and I’m refining this, so genuinely appreciate the challenge.
If you’re open to it, I’d actually value your feedback on the real flow , happy to share access so you can test how it behaves end-to-end.
That’s a fair concern ,and I agree most people won’t put their deepest thoughts into a random app.
A lot of people are already using tools like ChatGPT for emotional conversations. But I think that’s also where a different issue is starting to show up.
Long, ongoing 1:1 conversations with AI can sometimes turn into over-processing or even a kind of self-isolation loop. It works for some people, but it’s not always lightweight or healthy as a default behavior.
What I’m building is intentionally the opposite of that.
It’s not a conversation tool at all.
It’s a quick, anonymous “let-go” moment inspired by the work of Jill Bolte Taylor, 90 second rule:
you express what you feel
it gets processed briefly
and then it’s gone
No history, no back-and-forth, no long sessions.
So it’s less about sharing “deep secrets” and more about:
releasing a feeling quickly and moving on
Totally get if it doesn’t resonate right away , it’s more of an experience than a concept.
If you’re open to it, I’d genuinely value your perspective after trying it.
Happy to share access .
Just looking briefly at your app, the design could use a lot of help. That first impression is important. I wouldn’t use it just from the login page.
Once I start testing it in the process I get stuck on the ‘Pick any resonate image’ step. I don’t upload an image, just type text. I click the button and get no response.
It’s important to give the user feedback. Why is a button not working? What is required for me to move on? That sort of thing.
First things first, work on that design.
Those are my first impressions. Hope that helps a bit. Keep working at it! You got this!
Are you using a local (browser inferred) model for minimization? Asking cause I understand PII can somewhat be detected and removed using code but I haven’t come across any code that can achieve complex abstraction.
I’m asking cause I am working on similar features and the only solution (to keep things fast and responsive) that I have found so far is running inference clientside with edge models like Gemma.
really interesting concept.
Keep working at it! You got this!
