Grade 1 Smart Tutor

Grade 1 Smart Tutor - How hard teaching a Grad 1 can be?

Comment

Inspiration

I was testing out standard AI learning tools, assuming that because these models are trained on so much data, they'd easily handle the basics. I typed in the simplest question possible: 'What is 1 + 1?' > The response was mind-blowing—but not in a good way. Instead of just saying '2', it gave me a deep dive into the famous 300-page mathematical proof of addition, explained binary math, and even discussed the concept of synergy. It was a brilliant answer for a university student. It immediately struck me: what happens when a young kid asks that exact same question? They'd be completely lost.

That was the spark for this project. We don't just need a smart AI; we need an age-appropriate AI that knows its audience and sticks strictly to the Grade 1 syllabus.

What it does

Grade 1 Smart Tutor is an AI-powered learning companion that helps students explore Math and English for now Curriculum-Grounded: It uses RAG (Retrieval-Augmented Generation) to answer questions strictly based on Grade 1 NCERT textbooks. Multimodal Learning: It doesn't just talk; it generates custom illustrations to explain concepts like addition or nature. Live Mode: A hands-free, real-time voice interface where children can speak naturally to the tutor, making learning feel like a conversation with a friend.

How I built it

Frontend: Built with React and Tailwind CSS, focusing on a high-contrast, kid-friendly UI. Reasoning & RAG: We used Gemini 3 Flash integrated with Vertex AI Search to ground the AI in a datastore of NCERT PDFs. Voice & Vision: Gemini 2.5 Flash TTS for high-quality, cheerful speech generation. Gemini 3.1 Flash Image for generating educational illustrations on the fly. Real-time Interaction: Implemented the Gemini Live API via WebSockets for sub-second latency voice conversations.

Challenges I ran into

I found that the Gemini TTS API returns raw PCM data (just a stream of numbers), which web browsers cannot play directly. I had to build a custom pcmToWav utility to manually inject a 44-byte "RIFF" header into the audio data so the browser would recognize it as a playable sound file.

Accomplishments that I am proud of

low-Hallucination Tutoring: strictly tune the system Instruction and implement Vertex AI Search grounding. We forced the model to say "I don't know" if the answer wasn't in the specific datastore, ensuring the tutor stays safe and curriculum-accurate. Seamless Live Mode: Achieving a natural conversation flow where the AI can be interrupted and respond instantly. Kid-Centric UX: A UI that balances simplicity with powerful multimodal features (Text + Voice + Image).

What I learned

Vertex AI Search is Good. It can find the exact details from were the data is fetched the Source - Like Joyful Mathematics (Class 1), Chapter 1: Finding the Furry Friends (Counting 1 to 5). which is fantastic that i can trust and its not making it up!.

What's next for Grade 1 Smart Tutor

Future of Smart Tutor can be Endless. From expanding it for all the Grades. providing Grade level Tutoring. Multi lingual implementation Guided learning under specific Syllabus stream lining for the curriculum / Schools Getting feedback on student understanding and changing the way of tutoring from the feed back.

Built With

Share this project:

Updates