Why we made our AI tutor voice-first (and what changed when we did)

Most AI study tools are a chat box. A student types a question, gets a beautifully formatted answer, copies it into their homework, and learns nothing. The tool did its job. The student didn't.

Real tutoring doesn't look like that. A tutor asks you to explain your reasoning out loud. Interrupts when you go wrong at step two instead of letting you finish a wrong page. Scribbles a diagram while talking. Notices you've hesitated.

That's the gap Kicci is built for. It's an AI voice tutor for Singapore students (Primary through A-Level, aligned to the MOE syllabus) — and increasingly, a tool for the tutors teaching them.

Why voice, and why this one

It's a conversation, not a transcript. Real-time bidirectional voice, so the tutor talks and the student talks back. You can cut in mid-sentence with "wait, why did you square both sides?" and it handles it — the way a person would, not the way a chatbot does.

It draws while it explains. Maths taught as a wall of text is not maths taught. Kicci generates diagrams and animated walkthroughs on the fly and narrates them as it goes.

It's grounded in the actual syllabus. Curriculum content sits behind a retrieval layer, so explanations match what's examinable in Singapore — not a generic internet average of how maths is taught somewhere else.

It's built for both sides of the desk. Students can learn alone. But tutors were the ones who kept asking for it, so most of what we've shipped this year is for them.

How it works — for students

Pick a teacher persona, pick a topic, start talking. Kicci teaches the concept out loud with visuals, stops to ask questions, and adapts to where you're weak. Worksheets are generated around your specific gaps and band level, with hints instead of answers — and you can submit answers one at a time or all at once for feedback.

How it works — for tutors

Three things, all designed around "the parts of teaching you'd happily delegate":

1. Live lessons. You design the lesson — objectives, sequence, the examples you want used. Kicci's voice agent delivers it. Your teaching, your structure, without you having to repeat the same explanation for the eleventh time this week.

2. Worksheets that grade themselves. Upload your materials or write a prompt. It generates questions (with diagrams where needed), students answer — typed or as a photo of handwritten working — and it marks them and explains the mistakes.

3. LiveQA. For anyone teaching live. Push multi-step questions to the class one at a time, students submit, AI grades every answer in real time, and you see the whole room's responses as they land — so you can jump on the common mistake immediately, or surface one student's good answer as the class example. Runs as a Zoom plugin and a Google Meet add-on, and works for tutors livestreaming on TikTok or YouTube.

Grading stops being the thing you do at 11pm.

What's working so far

Kicci is officially going live soon — 300+ students in waiting list , the first tuition centre onboarded, and tutors currently testing the teaching-assistant flow. One of our maths explainer videos recently hit ~500K views, which suggests the demand for "explain this properly" is not exactly small.

What I'd like this forum's take on

For the teachers and tutors here: which part of teaching would you actually hand to an AI, and which part would you never? My instinct is that grading and drilling go first and lesson delivery goes last — but the tutors I've spoken to keep surprising me on this.

And for anyone who's built voice AI products: how are you handling the moment when the model is confidently wrong out loud? Text lets a user re-read and catch it. Voice doesn't.

Kicci is at kicci.study — happy to answer anything about the stack, the pedagogy, or the parts that broke.