Mira
A language partner, inside a chat
Ask about a kanji and it draws itself, stroke by stroke, right in the conversation. A chat app I'm building to learn languages the way they're learnt: by talking.
Watch it, then trace it.
A language is learnt by talking, so Mira isn't a flashcard deck. It's a chat, and the kanji turn up inside it. One real recording, start to finish. Scroll to play it, scroll back to rewind:
Ask the way you'd ask anyone.
"How do you write forest in Japanese?" The answer streams in as plain talk: 森 is three trees stacked, and each 木 is written the same way, trunk first, then the branches.
Then it draws itself.
The stroke screen opens on its own and builds 森 in order, all twelve strokes: the top tree, then the two below.
Then you trace it.
Your turn, stroke by stroke. Get one wrong and it shows you where it should have gone. The hints fade as you get better.
The chat takes its name from it.
Close the drawing and you're back in the conversation, which is now titled 森, so it's easy to find again later.
Tap it, and it draws again.
The reply keeps a faint underline under every kanji, here 森 and 木. Tap one, any time, and it draws itself again.
Eleven agents became one
The first version had eleven specialists: career, music, dreams, writing, language practice and more. A router sent each message to one of them, each with its own thread and memory. It worked, and it felt wrong. You could see the seams: change the subject and you met a different personality.
So the specialists became tools one assistant can consult. The model does the routing now, and memory lives with the assistant instead of in each chat.
The cost is control. Routing used to be code I could read and test; now it's the model's judgment, and the traces are how I see what it decided.
Talk to it, and the kanji come to you.
Speech in, speech out: say a word mid-conversation and its character draws itself on screen while Mira answers. After that, a partner for spoken French.
Tap a card to see it drawn