LearniGPT — Adaptive micro-lessons that fix what you forget

Home Screen
Count increasing when clicked (INTERACTIVE)

Inspiration

Every student forgets — but what if AI could remember what you forget for you? We were inspired by watching classmates repeatedly struggle with the same concepts even after multiple explanations. Teachers often don’t have time to personalize feedback for each mistake. We wanted to build an AI system that identifies those misconceptions instantly and delivers short, targeted lessons that actually fix them.

What it does

LearniGPT is an AI-powered micro-learning platform that personalises education in real time. It detects what a student misunderstands from quiz answers and instantly generates a 2–5 minute micro-lesson — complete with a clear explanation, a worked example, and tailored practice questions.

Key features: Misconception Detection: Classifies incorrect answers into common misunderstanding patterns. AI Micro-Lesson Generator: Uses LLMs to generate mini lessons adapted to learning style and reading level. Spaced Recall Scheduling: Reminds students at optimal intervals for memory retention. Teacher Dashboard: Summarises class misconceptions and allows assigning AI-generated micro-lessons. Accessibility: Read-aloud and simplified-language modes for diverse learners. LearniGPT makes personalised learning scalable, accessible, and truly adaptive.

How we built it

We built LearniGPT as a React + Tailwind web app powered by a FastAPI (Python) backend. Frontend: React for interactive quiz-taking, micro-lesson playback, and teacher dashboards. Backend: FastAPI connects to the AI models and spaced-recall logic. AI Core: A lightweight scikit-learn classifier detects misconceptions from quiz data. An LLM (prompt-engineered) generates concise, structured explanations, examples, and practice tasks. Database: SQLite for demo storage (students, quiz results, micro-lessons). Accessibility: Integrated browser-based text-to-speech (Web Speech API). We deployed the app using Vercel (frontend) and Render (backend).

Challenges we ran into

Balancing creativity and control in LLM outputs: We had to design strong prompt templates to keep micro-lessons concise and educationally accurate. Generating realistic synthetic data: Building believable student mistakes for the classifier was harder than expected. Time management: Integrating the full quiz → classifier → generator → scheduler pipeline in two weeks was ambitious. Accessibility testing: Ensuring voice readouts and simplified language were smooth on multiple browsers took trial and error.

Accomplishments that we're proud of

Built a working end-to-end adaptive tutoring prototype in under two weeks. Achieved consistent, personalised micro-lessons using prompt engineering — lessons felt human, not robotic. Integrated AI explainability: each generated lesson lists the “detected misconception,” so learners know what they misunderstood. Added text-to-speech and teacher analytics dashboard for inclusivity and real-world relevance. Received strong feedback from beta testers: “It felt like the AI actually understood where I went wrong.”

What we learned

How to combine classification + generation models for a powerful educational loop. The importance of pedagogical framing — AI isn’t useful unless its output supports learning principles like retrieval practice and feedback. How prompt wording drastically changes educational quality (we tested over 30 variations). Accessibility isn’t an afterthought — inclusive design made LearniGPT better for everyone.

What's next for LearniGPT — Adaptive micro-lessons that fix what you forget

We plan to: Expand to multiple subjects — math, science, and language learning. Add adaptive difficulty curves and emotion-aware tutoring (detect frustration via response time and tone). Build teacher assignment tools to auto-create homework from misconception analytics. Collaborate with schools to pilot LearniGPT for real classrooms using anonymised student data. Open-source the micro-lesson generation framework so other educators can build AI-powered learning modules. Our mission: make forgetting impossible — one adaptive micro-lesson at a time.