Equity.edu

Inspiration

STEM students often get stuck in the middle of a problem and do not know exactly where their reasoning broke down. Traditional chatbots can answer questions, but they usually cannot see the student’s written work, follow the evolving notebook context, or respond in a way that feels like a live tutor inside the workspace. We wanted to build a more empowering kind of support: a free, 24/7 personal tutor that lives directly inside the whiteboard students are already using. By combining real-time workspace awareness with multilingual support, Equity.edu makes help more accessible, immediate, and equitable, so students can get guidance in the same place they think, write, revise, and learn.

What it does

EduEquity is a GoodNotes-style AI whiteboard for STEM tutoring. Students can upload class materials, write directly on the board, ask questions out loud, and get contextual help without leaving the page.

Core experience

Students write with mouse, trackpad, or stylus on a tldraw canvas.
They upload PDFs or images of worksheets, notes, or formula sheets.
Uploaded files can be used as tutor context only, or placed directly onto the whiteboard so students can write on top of them like a digital notebook.
The student asks a question out loud, like “Is this right?”, “Show me how to do integration by parts,” or “Create a visual for this.”
The app locks the board during processing, captures the current whiteboard state, and sends the conversation plus canvas context to the tutor pipeline.
The tutor can respond in multiple ways:
- a spoken Socratic hint
- a targeted annotation/highlight over the student’s work
- a practice problem written onto the board
- a generated visual explanation drawn directly on the whiteboard
Voice output is spoken with ElevenLabs, and the board remains locked until playback finishes.

Beyond basic tutoring, EduEquity also supports

multilingual tutoring, including language-aware speech and UI changes
rolling conversation memory so the tutor can follow short replies like “yes” or “I got 12”
board-aware visual generation for graphs, concept diagrams, and worked demonstrations
a live session summary / progress letter view based on the conversation and course materials

How we built it

Next.js App Router + TypeScript + Tailwind
tldraw as the notebook / whiteboard layer
custom board locking and screenshot capture during AI analysis
/api/analyze for routing tutor requests across:
- local AMD-hosted vLLM models (Qwen2.5-7B-Instruct and Qwen2.5-VL-7B-Instruct) for text + vision reasoning
- Groq llama-3.3-70b-versatile as a fallback / structured visual-planning path
- OpenRouter-configured Gemini Flash fallback where needed for vision support
/api/speak for ElevenLabs streaming TTS using eleven_turbo_v2_5
uploaded PDF/image extraction plus board rendering so materials can be both read by the tutor and displayed on the canvas
conversation history tracking so the tutor can respond naturally across multiple turns

Key features

Whiteboard-native tutoring: The AI responds inside the same workspace where the student is writing, instead of in a detached chat box.
File-grounded tutoring: Students can upload worksheets, notes, or formula sheets, and the tutor bases its guidance on those materials.
On-board PDF workflow: Uploaded PDFs can appear directly on the whiteboard, where students can annotate them like a digital notebook.
Visual explanation mode: When a student asks to “show me” something, the AI can draw diagrams, graphs, step flows, and worked visual explanations below the existing board content.
Conversational memory: The tutor keeps track of recent dialogue, so follow-up responses feel like a real exchange rather than isolated prompts.
Multilingual support: The tutor can respond in the student’s selected language, including spoken output and translated interface text.

Challenges

Keeping whiteboard interaction stable while layering AI features on top of tldraw
Preserving coordinate alignment for annotations and generated visuals after uploads, panning, and zooming
Avoiding overlap between uploaded worksheet pages and AI-generated board content
Making voice playback control the board lock state reliably, without using fragile timers
Constraining model outputs into structured JSON for annotations, visuals, and tutor actions
Managing PDF rendering and board performance on touch devices like iPad