Polyglot: A Clean-Architecture, LLM-Powered Platform for Interactive Language Learning

Learners in emerging markets often lack affordable access to native-speaker practice, pronunciation feedback and writing guidance. This paper describes Polyglot, a full-stack language-learning platform built during an internship at the AMOND Association (Rabat). A React 19 single-page application communicates with an asynchronous FastAPI backend that persists data in SQLite through SQLAlchemy 2.0. Ten language services run on the Groq-hosted LLaMA 3.3 70B model through structured prompts and a defensive JSON-extraction layer, without fine-tuning. The platform parses video lesson transcripts from public sources, offers speaking practice and grammar analysis, schedules vocabulary review with the SM-2 spaced-repetition algorithm, and provides analytics and administration views behind JWT authentication. Both tiers follow a four-layer Clean Architecture. We evaluate the system against its functional and non-functional requirements using manual testing and report six known limitations, including the absence of an automated test suite, SQLite as the production store, hallucination risk and partial rate limiting. We present Polyglot as a working prototype, not a validated pedagogical intervention: no learning-outcome study was conducted. Preprint, not peer reviewed.