Language learning and real-time communication platform with LLM-driven adaptive grammar support: A review and proposed framework

Abstract

English learners often lack two things at once: regular conversation with real partners and grammar practice that matches their individual weaknesses. This paper reviews recent LLM-based and adaptive English-learning systems and proposes an integrated web platform that combines friend-gated language-partner discovery, real-time chat and audio/video calls, and an LLM-driven adaptive grammar module. The platform uses React.js, Node.js with Express.js, MongoDB and the Stream SDK. The adaptive module analyses learner text, categorises grammatical errors, maintains an error-weighted learner profile through an exponential-moving-average update, and selects the next exercise by maximising a utility function that balances weakness relevance, difficulty fit and repetition. A reciprocal compatibility score ranks candidate language partners. The paper formalises these components mathematically, presents architecture, workflow, sequence and data-model diagrams, and defines an evaluation protocol that compares adaptive practice with a fixed-sequence baseline. No experimental results are claimed: the numerical illustration uses synthetic inputs only, and implementation and empirical testing are required to establish whether the approach improves learning outcomes.