AI can teach, evaluate and support students with minimal cost and high precision

In a bold step toward democratizing access to cutting-edge learning tools, researchers have developed a powerful, open-source artificial intelligence (AI) framework capable of transforming traditional university courses into highly interactive, intelligent digital environments. This new AI-based educational platform enables instructors with no technical expertise to deploy course materials enhanced by large language models (LLMs), providing students with 24/7 access to an AI assistant and evaluator.

The study, titled “Enhancing the Learning Experience with AI” and published in Information (May 2025), evaluates the platform’s usability, technical performance, and impact across four pilot university-level STEM courses, reporting near-human or superior performance in automated teaching assistance and student evaluation.

How does the AI teaching assistant framework work?

The proposed system is modular and structured into six core components - four student-facing and two instructor-facing. Students interact with an AI-powered assistant (Module 1) capable of answering questions about course material by retrieving contextually relevant excerpts from uploaded PDFs or documents. Additional modules support mock exams (Module 2), final assessments (Module 3), and feedback collection (Module 4).

Instructors, meanwhile, use a simple drag-and-drop interface (Module 5) to upload lecture content and generate or import test questions. A statistical module (Module 6) helps track year-over-year performance, comparing AI-enhanced outcomes to those from traditional teaching methods.

Under the hood, the platform uses Retrieval-Augmented Generation (RAG) with state-of-the-art LLMs like Gemini 2.0 and Claude Sonnet. The AI assistant augments student questions with relevant content snippets before generating an answer. The system is entirely deployable via Google Cloud Run and costs less than USD 2 per course per year using efficient models like Gemini 2.0 Flash or DeepSeek.

What were the real-world results from pilot studies?

The platform was tested across four university courses, Databases, Database Programming Techniques, Object-Oriented Programming, and Designing Algorithms, at the Constantin Brancusi University in Romania. These courses involved more than 150 students and were evaluated using both traditional and AI-assisted formats.

Students using the AI-enhanced platform received real-time feedback, self-assessment tools, and adaptive question generation. Evaluations were conducted through manual review by instructors, automated LLM scoring, and benchmarks using RAG-specific libraries like Ragas.

Performance metrics showed that:

• Single-choice answer correctness reached 100%.
• Free-form answer accuracy ranged from 95% to 100%, surpassing human expert benchmarks.
• Relevancy and faithfulness metrics remained consistently above 85% for the top LLMs used.
• Evaluation modules graded with 90–100% accuracy, offering actionable, personalized feedback to incorrect answers.

Students praised the assistant’s 24/7 availability, clarity of answers, and the ability to ask questions without fear of judgment. Instructors highlighted ease of setup, effective AI grading, and a dramatic reduction in repetitive query handling, enabling them to focus on complex student issues.

What barriers does the platform address and what’s next?

Despite its success, broader adoption of AI in education faces hurdles: insufficient technical training among teachers, limited institutional budgets, and regulatory ambiguity around AI use in classrooms. The platform’s authors address these by offering a free, open-source, browser-based tool that runs on minimal hardware and supports multilanguage content.

Cost assessments reveal that the system can run at approximately USD 1–2 per class annually using top-performing LLMs. Unlike expensive MOOC platforms such as Coursera or edX, which charge per course or by subscription, the solution offers unrestricted institutional deployment at negligible operational cost. Its decentralized model and customizable privacy settings make it particularly attractive for public institutions concerned with data sovereignty and confidentiality.

In terms of policy, while regulatory frameworks around AI in education remain nascent, the authors argue that since students are already enrolled in official academic programs, the framework introduces no new ethical challenges. It merely augments pre-approved content with intelligent assistance and grading tools.