PolEdu

March 2026. Personalized AI math tutoring platform that adapts lessons to each learner's profile. Built for LotusHacks 2026 — 5,000 applicants, 800 participants across 220 teams. Didn't place.

the problem

One-size-fits-all tutoring doesn't work. A hands-on learner and an auditory learner sitting through the same lesson will have completely different outcomes — but most edtech products ignore this entirely. The content is identical; only the pacing changes.

The result: learners disengage, gaps compound, and the platform gets blamed for the student's learning style.

what PolEdu does

PolEdu profiles each learner on intake and uses that profile to shape every lesson it generates — not just the difficulty, but the format of explanation.

Adaptive content by learning style:

• Hands-on learners get interactive sliders, parametric graph explorers, and step-by-step problem reveals
• Auditory/reading learners get analogy cards and text-to-speech narration
• Quizzes adapt to the same profile — 10 questions across multiple-choice, numeric, and true/false formats

AI-generated lessons via chat — lessons are delivered conversationally, with structured content blocks rendered inline: static graphs, interactive visualizations, KaTeX-formatted equations, and real-world analogy cards.

IELTS mock test generation — full four-section mock tests generated on demand, with source material retrieved via Exa and stored in ChromaDB for consistent retrieval.

architecture

Frontend: Vue 3 + TypeScript (Vite)

Backend: FastAPI (Python)

AI layer:

• Google Gemini (Gemma 3 12B) for chat routing and intent classification
• OpenAI GPT for lesson and test content generation
• Exa API for web research on source material
• ChromaDB for vector-based document retrieval

Lesson content blocks include: text, static graphs, interactive graph explorers, parametric graphs with sliders, step-by-step reveals, analogy cards, and KaTeX-rendered math throughout.

demo personas

Two demo profiles ship with the app — Alex (hands-on learner) and Jamie (auditory learner) — so judges and reviewers can immediately see how the same topic renders differently depending on the learner type.

reflection

Didn't place. My teammates oversold their skills on intake and underdelivered throughout — I ended up building most of it alone.

The bigger lesson though: AI makes everyone fast now. Every team at a 220-team hackathon has Cursor open and Claude on standby. Solo-or-effectively-solo doesn't cut it when the baseline execution speed has been equalised. What actually separates teams at this level is originality — a sharper insight into the problem, a more surprising angle, something the judges haven't seen the fifth variation of that weekend. PolEdu was well-built but not surprising. That's on me.