Abstract
<p>Large language models are now routinely used to summarize lectures and generate study material, butmost published systems are built and tested on tidy, single-language English audio in well-connectedsettings. That leaves out a lot of students. Across much of undergraduate education outside North Americaand Western Europe, instructors move between English and a home language mid-sentence, and internetaccess in the classroom or hostel is not something you can count on. This paper describes a browser-basedlecture companion built for those conditions instead. It records live lectures in English, Hindi, or the HindiEnglish mixing (Hinglish) that Indian classrooms actually sound like, transcribes them in the browser, anduses a language model to turn the transcript into structured notes, flashcards, flagged learning gaps, andcalendar reminders. Everything is stored locally first, so the tool keeps working when the connectiondoesn't. Working from the Design Science Research tradition in information systems, we treat thedeployed application itself as the research contribution: we document its architecture, read its ten studymodules as an implementation of Zimmerman's cyclical model of self-regulated learning, and map eachdesign decision onto a specific, cited gap in the literature on LLM-assisted note-taking, code-switchingNLP, active recall, learning-analytics dashboards, intelligent tutoring systems, AI reliability, the digitaldivide, and multilingual education policy. We also report a structured heuristic usability audit againstNielsen's ten heuristics and a feature-level comparison against three widely used tools — Otter.ai, GoogleNotebookLM, and Anki — and lay out a pre-registered, quasi-experimental protocol for a learningoutcomes study we have not yet run: hypotheses, instruments, a statistical analysis plan, and a fullaccounting of threats to validity. Throughout, we try to keep a clean line between what the paper actuallyshows (a working artifact, its design logic, an expert audit, a market comparison) and what it only proposesto test later (whether any of this improves retention or self-regulated learning). That separation betweenbuilding and evaluating is the whole point of the DSR framing we're borrowing.</p>