• 

• GIF

• 

Inspiration

The current digital education paradigm is broken in two major ways. First, text-based LLMs are creating an "AI copy-paste" crisis—students generate perfect text answers without understanding the underlying concepts (rote learning or ratafication). Second, high-fidelity visual platforms rely on heavy video streaming, which constantly buffers on slow networks and consumes massive bandwidth, excluding millions of students in low-to-mid bandwidth regions. We were inspired to build an AI tutor that forces conceptual learning without the data tax.

What it does

My Magic Pencil is an AI-powered tutor that replicates the experience of a live teacher at a digital blackboard. Instead of outputting text blocks or streaming video, it translates natural language prompts into dynamic, animated mathematical drawings paired with vernacular voice synthesis (supporting 88+ languages). It explains complex STEM concepts step-by-step, stroke-by-stroke, enforcing conceptual engagement and preventing copy-pasting.

How we built it

We engineered a proprietary Vector-Agent Engine. Instead of treating AI output as static text or pixels, we orchestrated a multi-agent system to infer spatial and mathematical logic from the user's prompt.

The Translation Layer: LLMs parse the educational request and break it down into logical steps.

The Math-as-Code Generator: The engine synthesizes this logic into dynamically animated Scalable Vector Graphics (SVGs) and coordinates it with TTS (Text-to-Speech) APIs.

Edge Rendering: We offload the visual rendering entirely to the user's edge device.

Challenges we ran into

Hallucinations in Spatial Logic: Standard LLMs are notoriously bad at math and spatial positioning. Getting an AI to reliably map coordinates for complex physics diagrams or geometric proofs required strict architectural guardrails and custom fine-tuning to force the model to output strict SVG syntax.

Audio-Visual Sync: Synchronizing the stroke-by-stroke SVG drawing animation with the multi-lingual voice output in real-time, especially considering network latency, was a major engineering hurdle.

#Accomplishments that we're proud of

90% Bandwidth Reduction: By transmitting "Math-as-Code" instead of video frames, we reduced server AWS costs and user data consumption by 90%, enabling 4K-quality visual pedagogy on 2G-level connections.

Proven Impact: We successfully deployed a live MVP pilot with over 500 engineering students at VIT Vellore, measuring a 26% boost in conceptual retention compared to standard text-based chatbots.

What's next for My Magic Pencil

We are currently scaling our B2B SaaS API to integrate our low-bandwidth "Blackboard" engine into global LMS platforms. We are also finalizing the global patent filings for our Math-as-Code architecture and expanding our vernacular language models for deeper rural penetration.

Built With

• amazon-web-services
• api
• edge-computing
• fastapi
• google-cloud-tts
• langchain
• llm
• openai-api
• python
• react
• svg