A local-first, AI-powered language-learning platform — structured lesson generation, an embedded tutor, role-play chat practice, and cross-device sync, built from scratch with no frontend framework.
Live demo: https://linguini-16g.pages.dev — try the sample lessons with zero setup, or add any LLM API key (or a local model) in the Models tab to generate your own.
Most language apps flatten everything to one level: grammar gets oversimplified, pronunciation becomes "sounds like...", and IPA, morphology, and glossing are off the table. Linguini was built to be the opposite — a maximally customizable AI learning tool that meets you where you are. A teaching-style switch runs every generation in either linguistic mode (IPA, morphological breakdowns, precise terminology for those with the background) or plain mode (no jargon, simple analogies for those without). Combined with free-form plan prompts, per-lesson refinement notes, editable everything, and swappable models per task, it produces real lesson plans that don't dumb down the content unless you ask them to.
- Plan generation — one prompt ("Advanced Korean for TOPIK II") becomes a 20–30 lesson curriculum with per-lesson grammar concepts and starter vocab.
- Full lesson pages — each lesson expands on demand into grammar sections with inline examples and translations, a vocabulary table with IPA, reference alphabet charts for phonetic lessons (Hangul, hiragana, Cyrillic...), and exercises marked by an AI checker.
- Practice — every lesson ships a role-play scenario that seeds a chat with a character bot; each user message is answered in character and run through a separate correction model.
- Tutor — a per-lesson Q&A chat embedded in the page.
- Vocab bank — every word from every lesson, aggregated, searchable, linked back to its source.
- Multilingual UI — English, 日本語, 繁體中文, 简体中文, 한국어, Español. The UI language also steers generation: explanations and corrections are written in it while the target language stays untouched.
- Accounts & sync (optional) — sign in to sync all state across web and Android; skip it and everything stays on-device.
Five independent LLM routes — plan outline, lesson body, quiz/scenario, role-play chat, error checking — each pointing at any provider: Claude, OpenAI, Gemini, DeepSeek, Kimi, Z.ai (GLM), OpenRouter, or a local LM Studio / Ollama instance. Heavy lifting can go to a strong model while per-message correction runs on something small and fast.
- Zero-framework frontend. Vanilla JS/HTML/CSS, no build step, no runtime dependencies. State → render functions → DOM, done by hand.
- Whole backend in one stdlib file.
server.py(pure Python, no pip installs) serves the app, persists state, and proxies LLM calls so browser CORS never gets in the way. - Deliberately simple storage. The entire app state is one JSON document in SQLite, with a human-readable JSON mirror that doubles as an automatic backup/restore path. Single-user app; a relational schema would buy nothing.
- Hosted flavor. The same frontend deploys as static files on Cloudflare Pages, with a single Pages Function replicating the LLM proxy.
- Sync done small. Supabase email auth + one row per user guarded by row-level security; last-write-wins with debounced pushes and pull-on-focus. No sync framework.
- Provider abstraction in code, not a library. OpenAI-compatible wire format by default, with in-code translation to Anthropic's Messages API (system promotion, auth-header style, response shape).
- Resilient generation. A tolerant JSON repair pass recovers truncated LLM output — partial-but-valid lessons instead of hard errors.
- Prompt-enforced pedagogy. A per-language script policy (native script vs. romanization by learner level) and a UI-language directive are injected into every prompt builder.
- i18n at 310 keys × 6 locales, dictionary-driven with scripted completeness checks.
- Android. A WebView wrapper bundles the same web files into an APK, with a small in-process Java HTTP server standing in for
server.py(persistence via SharedPreferences, same LLM proxy contract).
- docs/ARCHITECTURE.md — runtime targets, the state→render model, the LLM pipeline (routes, proxy inversion, wire-format translation, JSON repair), the prompt system, and i18n.
- docs/DATA-AND-SYNC.md — the state document and normalizers, the three persistence tiers, the Supabase sync protocol, and failure modes.
python3 server.py # http://127.0.0.1:5173 — Python stdlib only, nothing to install├── index.html / styles.css / app.js # the whole frontend (paper-notebook theme)
├── i18n.js # 6-locale dictionary + t()
├── samples.js # localized sample lessons (landing page)
├── sync.js # Supabase auth + cross-device sync
├── server.py # static + state + LLM proxy (stdlib)
├── functions/api/chat.js # Cloudflare Pages Function: hosted proxy
├── deploy.sh # stage + deploy to Cloudflare Pages
└── android/ # WebView wrapper + in-process Java server