Lizm.

It started as a post. 30k people watched the idea land on LinkedIn. This is what shipping it looks like.

The model router. One call site decides between device, cloud, or proxy for every request. Online, it upgrades to Claude; offline, it falls back to the device. Either way, there’s always an answer.

Gemma, on the device. Gemma 4 E2B looked better on paper, but it crashed an iPhone 14 outright: 6 GB of RAM isn’t enough once iOS takes its cut. Gemma 3 4B, quantized to 2.7 GB, fits, and keeps the guide talking with no signal at all.

BYOK: bring your own key. Thirty walk plans and a hundred voice questions come free every month, on my dime. Blow through that and drop in your own Claude API key: the shared quota steps aside and you’re on your own budget, not mine.

Spoken in any language. Gemma translates the narration locally before it’s read aloud, so it’s free, private, and fully offline. The cloud stays reserved for conversation, not bulk translation.

Private by default. Your conversations, profile, and saved plaques never leave your device. Only Claude calls go out, brokered through a single proxy; everything else stays put.

Context is king. Your active pack is the model’s context. Every reply weaves in the words you’re learning and stretches a little past them, so your vocabulary grows as you speak.

Works at 35,000 feet! A 668 MB Gemma 3 1B runs right in the browser tab, so the conversation keeps going with no signal. Your deck lives on your device and syncs safely to the cloud the moment you land.

Translation that reads the room. Tap a word and a context pass sends the whole sentence, not just the word, so ‘espèces’ files as cash, not species. Offline, a local model (mBART-50) keeps translating on your device: private, free, no round-trip.

From PDF to a fresh pack. Drop in a PDF and it’s parsed into cards, a whole pack in one go. Or snap a printed page and OCR lifts the words, straight from it. Paper and screens both become vocabulary.

Voices worth imitating. Google’s Chirp 3: HD voices, 54 of them across eight languages, read every line in native-quality speech, so you’re shadowing a real accent, not a robot.

Credits, in plain sight. Everyone starts with 100 credits a day, enough for a real session. Generous by default; Plus and Pro raise the ceiling.

Speech-to-speech pipeline. Hugging Face’s speech-to-speech runs as a local Python server with four models from the Hub: Parakeet for speech to text, Smart Turn for end-of-turn detection, Qwen3 4B as the language model and Qwen3-TTS for the voice. All running locally.

Voice and latency. Qwen3-TTS is slower than Kokoro, so we cache the audio for fixed game lines. After the first take, every command and score line plays instantly.

Game logic in plain JavaScript. The browser runs the game: rounds, timing, scoring and the nine commands. Game lines are sent as out-of-band responses, so they never grow the model’s conversation history.

Turn-taking. The mic stays open for the whole session. Smart Turn decides when you’ve finished speaking from how you stopped, not just from silence. If you speak while Simon is busy, his reply pauses rather than being dropped.

Webcam referee. MediaPipe pose tracking runs in the browser and compares each move to your starting stance. A move only counts if you hold it for a second, and 57 unit tests cover the rules, including mirrored left and right. Video never leaves your laptop.

Upstream fixes. Building on speech-to-speech surfaced three tool-call bugs, including one where Simon read his own tool calls out loud. We fixed all three, with tests, on a branch ready to send upstream.