Adaptive AI Tutor - An Intelligent PDF-Based Learning System Using Local Large Language Models

Abstract

This review paper examines the role of Artificial Intelligence (AI), Machine Learning (ML), and Deep Learning (DL) in enhancing recommendation systems for e-commerce platforms such as Amazon, Walmart, Alibaba, Flipkart, and Netflix. Recommendation systems help users discover products that match their preferences, improving customer satisfaction, engagement, and sales. Traditional recommendation methods, including collaborative filtering, content-based filtering, and hybrid approaches, often face challenges such as cold-start problems, data sparsity, and limited personalization. AI-driven techniques have emerged as effective solutions to overcome these limitations. The study reviews recent research from the last five years on intelligent recommendation systems (IRS). ML approaches such as supervised learning, reinforcement learning, matrix factorization, sentiment analysis, fuzzy logic, and optimization algorithms improve recommendation quality by analyzing customer behavior, ratings, reviews, and purchasing patterns. DL techniques, including CNNs, RNNs, LSTMs, GRUs, word embeddings, and attention mechanisms, further enhance recommendation accuracy by capturing complex user-product relationships and sequential behavior patterns. AI-powered recommendation systems utilize key components such as collaborative filtering, content-based filtering, hybrid models, sentiment analysis, image-based recommendations, graph neural networks, and reinforcement learning. These systems analyze browsing history, clickstream data, search queries, purchase history, reviews, and customer preferences to provide highly personalized product suggestions, discounts, offers, and advertisements. Advanced approaches also incorporate visual search, natural language processing (NLP), and social network information to improve recommendation relevance. The review highlights several practical applications of intelligent recommendation systems in e-commerce. Businesses use AI-driven recommendations to increase customer retention, conversion rates, click-through rates, and overall user satisfaction. Features such as dynamic content personalization, session-based recommendations, omnichannel integration, predictive cart abandonment detection, and context-aware suggestions allow companies to adapt recommendations in real time. Major e-commerce companies like Amazon, Walmart, Alibaba, and Flipkart rely heavily on AI-based recommendation engines to drive engagement and revenue growth. The findings indicate that AI-based recommendation systems significantly outperform traditional recommendation methods. Deep learning and hybrid recommendation models achieve recommendation accuracies exceeding 90%, while effectively addressing challenges such as cold-start issues, scalability limitations, and data sparsity. Techniques such as CNNs, LSTMs, BERT, sentiment analysis, and graph neural networks provide more accurate, adaptive, and customer-centric recommendations. The paper concludes that AI-powered recommendation systems are transforming modern e-commerce by delivering smarter personalization, better customer experiences, and higher business performance. Future research should focus on emerging technologies such as transformers, graph neural networks (GNNs), reinforcement learning, explainable AI, and privacy-preserving recommendation systems to further improve scalability, transparency, adaptability, and recommendation accuracy.

Introduction

This paper presents the Adaptive AI Tutor, an offline, AI-powered educational system designed to help students understand complex academic PDF documents using locally hosted Large Language Models (LLMs). Traditional learning platforms provide static content and often depend on cloud-based AI services, which raise concerns about privacy, internet dependency, and subscription costs. To address these issues, the proposed system uses the Ollama framework with local LLMs such as Mistral and Qwen, enabling fully offline and privacy-preserving learning.

The system automatically analyzes uploaded PDFs to identify difficult concepts, generates multimodal explanations (text, summaries, analogies, diagrams, and glossaries), creates adaptive quizzes based on learner performance and confusion levels, recommends educational videos and academic resources, and supports multiple languages including English, Hindi, and Telugu. It also stores user progress and session history, allowing learners to resume previous study sessions.

The architecture consists of four layers: a Streamlit-based user interface, an LLM intelligence engine, a SQLite database for user and learning history management, and a PDF processing module for text extraction and analysis. The methodology includes PDF parsing, AI-based confusion detection, adaptive quiz generation, resource recommendation, and multilingual content generation.

Experimental evaluation across various academic subjects demonstrated effective identification of difficult content, high-quality explanations, accurate quiz difficulty adaptation, and reliable session restoration. The system achieved strong performance in concept analysis and multilingual support while maintaining complete offline functionality.

Conclusion

This paper presented the Adaptive AI Tutor, a fully offline, locally hosted, intelligent PDF-based learning system powered by open-source large language models via the Ollama framework. The proposed system addresses critical gaps in existing educational AI tools by delivering adaptive difficulty assessment, multimodal explanation generation, curated resource recommendations, multi-language support, and persistent session management — all without cloud API reliance, ensuring complete data privacy and zero recurring operational cost. Experimental evaluation demonstrated the system\'s effectiveness in identifying conceptually challenging content, generating contextually relevant multimodal explanations, calibrating quiz difficulty to individual learner profiles, and maintaining reliable session persistence. The system establishes a practical and accessible foundation for AI-driven personalized tutoring in resource-constrained and privacy-sensitive educational environments.

References

[1] J. R. Anderson, C. F. Boyle, and G. Yost, \"The geometry tutor,\" in Proc. 9th IJCAI, 1985, pp. 1–7. [2] K. VanLehn et al., \"The Andes physics tutoring system: Lessons learned,\" Int. J. Artif. Intell. Educ., vol. 15, no. 3, pp. 147–204, 2005. [3] A. T. Corbett and J. R. Anderson, \"Knowledge tracing: Modeling the acquisition of procedural knowledge,\" User Modeling User-Adapted Interaction, vol. 4, no. 4, pp. 253–278, 1994. [4] C. Piech et al., \"Deep knowledge tracing,\" in Adv. Neural Inf. Process. Syst., vol. 28, 2015. [5] G. Kurdi et al., \"A systematic review of automatic question generation for educational purposes,\" Int. J. Artif. Intell. Educ., vol. 30, no. 1, pp. 121–204, 2020. [6] H. Touvron et al., \"LLaMA: Open and efficient foundation language models,\" arXiv:2302.13971, 2023. [7] A. Q. Jiang et al., \"Mistral 7B,\" arXiv:2310.06825, 2023. [8] J. Bai et al., \"Qwen technical report,\" arXiv:2309.16609, 2023. [9] Y. Liu et al., \"Text summarization with pretrained encoders,\" in Proc. EMNLP, 2019, pp. 3730–3740. [10] A. Aleven et al., \"Tutoring answer explanation fosters learning with understanding,\" in Proc. AIED, 1999, pp. 199–206. [11] M. J. Woolf, Building Intelligent Interactive Tutors. Morgan Kaufmann, 2008. [12] P. Graesser et al., \"Intelligent tutoring systems,\" in Technology-Enhanced Learning, Springer, 2018, pp. 249–265.

Copyright

Copyright © 2026 Kanimozhi M, Bakkamantula Vamsi, Chintha Ramesh Reddy, Chimata Pavan Kalyan. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.