Hold the button, mumble what you're weighing, snap the menu if there is one. WDYT works out the actual question and the options, shows them to you to fix, and only then gives you an answer — with a number for how sure it is.
iPhone · coming to the App StoreHold or tap the mic and talk. English through Apple's recogniser; Chinese through SenseVoice running on your phone. No audio ever leaves the device.
A generator model reads your words and photos and writes the question, the options, and what it saw — you never fill in a form.
Nothing is decided until you've looked. Edit the question, delete an option, add one it missed, correct what it read off the photo. Then confirm.
The answer comes from TypeSafe's Jev, a model that returns a typed choice with calibrated probabilities in under half a second — no rambling.
When it's 55/45, it says so. When you gave it nothing to go on, it says that too, and lets you add one sentence and ask again.
Two plain files — identity.md and soul.md — hold who you are and how you decide. Off by default; when on, they're revised after each answer and you can edit them any time.
Understand → confirm → decide. The decision model is only called after you confirm.
“What should I eat…” plus a photo of the menu. Up to three photos; they're resized on the phone before anything is sent.
The question in one sentence, the options with a line each, and “what I saw” from your photos. Fix anything. Nothing has been decided yet.
One pick, the probability of every option, a confidence figure, and one line on why. Add a sentence and ask again if it missed something.
WDYT has no backend. Speech is recognised on the device. Understanding goes straight from your phone to Google Gemini with your own API key; the decision goes straight to OpenRouter (TypeSafe Jev) with your own key. Keys live in the iOS Keychain. Privacy policy · Terms
Two different jobs. Gemini (Google) reads your words and photos and writes the question — you need a Gemini key from Google AI Studio. Jev (TypeSafe) makes the decision — it's reached through OpenRouter, so you need an OpenRouter key. Both are pay-as-you-go on your own account; a typical question costs well under a cent. WDYT never sees either key.
Jev is a “System One” model from TypeSafe: it doesn't generate text, it returns a typed answer — one option from your list, a yes/no probability, or a position on a scale — with calibrated probabilities, in about 100–500 ms. That's what makes the confidence figure meaningful and the answer instant. Gemini is used for the part Jev can't do: understanding fuzzy speech and reading photos.
No — Jev is text-only. Gemini writes down everything it saw (every menu item with its price, what the outfit looks like) and that text is what Jev judges. You see that text on the confirm screen and can correct it.
The interface is English, 繁體中文 or 简体中文 — your choice, independent of the phone's language. Speech: English via Apple; Mandarin via SenseVoice after a one-time ~230 MB download. The model is shared with other tautiu.dev apps on your phone, so you only download it once.
identity.md is who you are — diet, budget, habits. soul.md is how you decide — what you optimise for, what you avoid. Both are sent with every question so answers fit you. You can write them yourself, or turn on “update these from my questions” and let Gemini revise them after each answer. Off by default.
Everything stays on your phone: keys in the Keychain, the two profile files, your history. The only network calls are the two you initiate, from your phone directly to Google and OpenRouter under their privacy policies.