lllanguage

Language learning from the gaps in real conversation – speak, catch what you needed, bring it back later.

Most language apps teach phrases you will never say. Real conversation is messier – you need a word mid-sentence, stay quiet, and forget to look it up later.

lllanguage is built around that gap. Speak with a partner, catch what you needed in the moment, bring it back when you have time. The vocabulary that actually showed up in your life.

Built with

Expo / React Native (audio streaming) · Fastify + WebSocket server · PostgreSQL · Local ASR / TTS / Ollama adapters · Next.js marketing + store pages

Where it got hard

01
Continuous PCM over WebSocket from a mobile client that cannot stream raw frames through Expo's stock recorder.
02
Realtime speak path: partial ASR, adaptive endpointing, barge-in, and stripping partner TTS echo from the learner's transcript.
03
A deterministic pedagogy reducer that owns durable conversation policy while the LLM proposes actions – so the model does not quietly rewrite the lesson.
04
A local Docker inference path so a pilot can run without burning a hosted model budget.

Role

Solo – design, code, and shipping.