Ariadot
The deadlines buried in your inbox, surfaced before they bit
I'm building Ariadot, an app that reads your personal Gmail and sends one short brief each morning of the things that slip through the cracks. An insurance renewal, a booking, a payment due next month, a form due before school starts. Too one-off for a task manager, annoying when you miss it.
The product was the easy part. What ate months was the trust problem: the whole pitch is "let an AI read your inbox," and nobody sane should be okay with that unless you can say, concretely, what the AI actually sees. "We take privacy seriously" is worthless. So the bar I set was simple. The model never sees a single real name, email, or account number.
How it works: before anything reaches a cloud model, every personal detail is swapped for a realistic fake. A name becomes a different name, an email becomes a different email. The model reasons on the fake version, and the real values are swapped back before anything is stored or shown. The model does the thinking and never learns who you are.
Three lessons that cost me the most, in case they save you time:
The small model that finds the names over-finds. It would decide "you", the verb "make", or the year "2026" was a person and mask it, which quietly corrupts the text. You need a dumb sanity-check layer on top of the smart one.
I added a second AI pass to "clean up" what the masking missed. Mistake. Ask a generative model to rewrite a document and it rewrites, it does not just swap. It merged two unrelated things into a company that never existed, and because the output gets saved and reused, the corruption compounded daily. Deleted it. The rule: find the names with AI, swap them with boring deterministic code that cannot get creative.
You can never prove masking is perfect, so assume you will miss one and build the thing that alerts you the moment you do, rather than claiming it is flawless.
Full writeup: ariadot.app/blog/masking-data-before-llm
It's iOS and Android, private beta, free, a handful of early users. I'd love honest feedback, especially the "I'd never trust an app with this" kind, and I'm happy to go deep on any of it. ariadot.app
2 Comments
This is a good way to frame trust: not “AI reads mail but trust us,” but “what crosses the boundary, and what is deterministic after that?” The failure mode you found with the second AI pass is the part I’d emphasize in the product: users need to know which steps are model judgment and which are boring code. For the “I’d never trust this” crowd, I’d want a visible local preview/audit log of the masked text before it leaves the device, plus a hard “exclude these senders/domains entirely” list. Email has too many weird edge cases (legal, medical, account recovery) for masking to be the only safety layer.
This is a strong trust problem to build around, and I think the masking writeup is probably more important than the app features right now.
The line I’d be careful with is “let an AI read your inbox.” Even if the privacy architecture is strong, that phrase makes the user picture the worst version of the product first.
The sharper frame might be:
“Get a morning brief from your inbox without exposing names, emails, or account details to the model.”
That makes the privacy mechanism part of the core promise, not a footnote.
I’d also make the beta onboarding less about asking users to trust the whole inbox product immediately, and more about proving one safe use case first: renewals, bookings, due dates, school forms, payment reminders. Those are concrete enough that people understand the value without feeling like they are handing over their whole life.
The product probably wins or loses on whether users understand exactly what the model does not see before they understand what the app can do.