PolEdu

Make learning fun again

Inspiration

We were inspired by our childhood experiences in learning things — specifically how a single good explanation, framed the right way, could make a concept click instantly. We wanted to build something that adapts to how a student learns, not just what they're learning.

What it does

PolEdu is a personalized English and Math tutoring platform. You tell it your name and how you learn best, and it generates a fully structured math lesson tailored to you — inline in chat, no page navigation required. Hands-on learners get interactive slider graphs and step-by-step guided problems. Listening learners get real-world analogies and narrated explanations. Every lesson comes with a 10-question mini-test with instant feedback. It can also generate full IELTS mock tests on demand.

How we built it

Frontend: Vue 3 + TypeScript (Vite), with custom canvas-based graph components and KaTeX for math rendering
Backend: FastAPI (Python) with a background job queue for async lesson generation
Lesson generation: GPT-5.2 with a structured JSON schema prompt, validated server-side across 3 retry attempts
Chat routing: Orchestrated intent classification with Dify with OpenAI GPT-4o as intent classifier, and Gemma 3 12B via the Gemini API as a fallback classifies every message into an intent (create lesson, ask topic, capabilities, etc.)
Research: Exa API fetches real web content on the math topic before generation
IELTS retrieval: ChromaDB local vector database over ingested PDF documents

Challenges we ran into

Getting GPT to reliably generate structured JSON lesson payloads with the correct block types for each learning style — required strict schema examples in the prompt and a validation + retry loop
Rendering math expressions and interactive graphs inside a scrollable chat window without overflow or layout breakage
Making the learning experience genuinely different between personas, not just cosmetically different — this required new block types (slider-graph, guided-steps, analogy) and per-section narration

Accomplishments that we're proud of

Two learners asking about the same topic get structurally different lessons, not just different wording
The inline lesson card renders fully inside the chat — interactive graphs, sliders, LaTeX, and a mini-test — without ever leaving the conversation
The backend validates every generated lesson against a strict schema before serving it, so malformed AI output never reaches the user

What we learned

Prompt engineering for structured output is as much about schema examples as it is about instructions — showing GPT the exact JSON shape you want is more reliable than describing it
Building adaptive UIs for different learning styles forces you to think deeply about pedagogy, not just product design
Async background jobs with polling are a practical pattern for expensive AI generation tasks that would otherwise timeout a request