The Method

Our conversational design is built on findings about cognitive load and language anxiety (Huang & Kakham, 2026). We correct sparingly rather than flagging every mistake, so feedback never overwhelms the conversation.

Not every mistake deserves the same treatment. Pronunciation slips get a gentle, in-flow correction; grammar and meaning errors get a clear, separate note — so nothing breaks your momentum, and nothing gets missed.

Research on AI conversation partners (CHI 2026) found that silent correction — quietly reformulating what you meant to say — often leaves learners, especially beginners, unsure what actually went wrong. So we don’t just recast: every correction comes with an expandable “why,” explained in your L1.

Research on feedback timing shows that delayed correction often produces stronger retention for meaning-based errors than interrupting mid-sentence. So instead of breaking your thought the moment you make a mistake, we let the exchange finish naturally, then surface the correction — protecting the flow that keeps you talking.

Generic AI tutors correct toward generic mistakes. We hand-built a taxonomy of the errors speakers of your L1 specifically make in French, English, and Spanish — gender agreement, word order (VSO vs. SVO), and more — so correction, and your progress score, reflects your actual starting point.

Most apps count how many mistakes you make. We track something better — whether a corrected form shows up correctly again later in conversation. That’s a stronger signal of real mastery than a raw error count, and it’s what moves you to the next stage.