Atlas: AI-Powered Learning Path Generator

Atlas is an educational tool designed to turn any collection of unstructured documents (PDFs, PPTs, DOCs) into a structured, highly optimized, and playable learning path. By leveraging the power of AMD GPUs and advanced Large Language Models, Atlas automates the extraction of educational concepts and intelligently deduces their prerequisite relationships. The core pipeline operates in six steps: First, raw documents are parsed and chunked based on semantic boundaries rather than fixed lengths. Next, local LLMs like Gemma, running efficiently on AMD GPUs via ROCm, extract raw learning concepts from these chunks. The system then clusters and deduplicates these concepts using cosine similarity and community detection algorithms. Prerequisite relationships are inferred by evaluating pairs with high similarity and asking Fireworks AI (Llama 3) to determine the learning direction. To ensure the resulting learning path is logically sound, the backend (built with FastAPI) runs cycle detection algorithms to eliminate contradictions, producing a valid Directed Acyclic Graph (DAG). Finally, topological sorting is applied to assign hierarchical tiers (foundational vs. advanced concepts). The final output is rendered via a modern Next.js frontend as a stunning, interactive skill tree. Learners can navigate this tree, taking checkpoint quizzes generated via RAG (Retrieval-Augmented Generation) directly linked back to the source documents. Atlas essentially builds a personalized, gamified curriculum from raw files.