AI-Powered Adaptive Learning Platform
An intelligent adaptive learning system that dynamically personalizes each student's experience using AI-driven question selection, auto-grading, and real-time skill tracking — not a static course platform.
Overview
Educational InstitutionsBuilt an adaptive learning system that relies fundamentally on AI to customize each student's learning experience dynamically. The platform doesn't serve static content — it understands the student's level in real time and selects the right question at the right moment using mathematical models and AI. The core engine combines Item Response Theory (IRT) with a 3-parameter logistic model for continuous student ability estimation, Bayesian tracking to follow student progress with every answer, and a multi-factor question selection algorithm that balances question difficulty, student performance, and question usage distribution to prevent repetition and bias.
What was not working
Traditional educational platforms serve the same content to every student regardless of their level. Manual question creation is slow and expensive. Grading is binary (right/wrong) with no meaningful feedback. There was no system that could simultaneously assess students in real time, generate questions at scale, provide intelligent feedback, and adapt difficulty dynamically — all while handling thousands of concurrent users.
What I built
Built a multi-layer AI system: 1. Content Generation Pipeline: Transformed question creation from a slow manual process to an automated pipeline — PDF → image conversion → visual mapping → Mathpix OCR → LLM parsing → review → Question Bank insertion. This enabled generating thousands of questions rapidly at a fraction of the cost. 2. Retrieval + Context-Aware Generation: Connected the model to the question database and subject-specific context (similar to RAG), operating not just on student input but also on their performance history and data. 3. Auto-Grading with LLMs: Built a non-binary grading system supporting partial credit and generating detailed feedback to help students understand their mistakes. 4. Prompt Engineering & Control: Designed different prompts per question type (MCQ / Essay / Math), controlled model outputs to reduce hallucination, and optimized outputs based on context (student level + subject + difficulty). 5. Performance & Scaling: Question selection runs in under 50ms, optimized retrieval from large question banks, Redis caching for latency reduction, and handling thousands of concurrent users.
Stack
07 partsWhat each part of the system does, from the screens the team uses to the services that send messages and reports.