BookNook is a multilingual reading platform built with SFU Linguistics for K-12 classrooms where students do not all share the same first language. Teachers upload one book, and students can read that same content in the language they are strongest in.

The platform itself was the research contribution. We built it as a seven-person team over five two-week sprints, and I worked as product owner while also contributing implementation work.

When a teacher uploads a book, it goes through a five-stage async pipeline — upload, OCR extraction (via PyMuPDF), a review gate where the teacher can fix any OCR mistakes, translation, then it's ready for students. Each stage can fail independently and the book just stops at that point instead of breaking everything downstream.

Translation runs through Meta's NLLB model, which supports over 200 languages including some that basically no classroom tech handles — Pashto, Somali, Amharic. The catch with NLLB is that loading it takes about 20 seconds. We cache it module-level as a singleton, so the first request pays the cost and everything after that is near-instant.

Each sentence gets rendered to audio via Edge TTS (fast, gives you word boundaries) with Matcha-TTS as a fallback (better quality for English). Word-level timing comes out of the TTS boundary events, which the frontend uses to drive karaoke-style highlighting as the audio plays. For languages where we can't get real boundaries, we fall back to a proportional estimate so the highlighting still works — it's not pixel-perfect but it's better than nothing.

After you finish a chapter, the app can generate a quiz. Questions come from the Groq API, prompted with child-focused instructions and broken into three difficulty tiers that map loosely onto Bloom's taxonomy — retrieval, comprehension, synthesis. The quiz system checks eligibility before letting you take it: have you read enough of the book, have you used up your attempts, are you still in a cooldown from the last one. That state machine lives in the backend so students can't cheat it by refreshing.

For teachers, there's an analytics dashboard that does sliding-window comparisons — your class's average score over the last 28 days versus the 28 days before that, per-student deltas so a teacher can spot who's trending down and check in before it becomes a problem.

The frontend is Next.js 14 with Tailwind. Accessibility was actually a real constraint given the user base — WCAG compliance, OpenDyslexic font as an option, high-contrast mode, dark mode, and PWA offline support so a student on a flaky connection can still finish a chapter they'd already loaded.

The repo totals about 40,000 lines across 446 commits from seven contributors, with 80% test coverage enforced in CI. Tests split across Pytest for the backend, Vitest for frontend units, and Playwright for end-to-end flows.