The itch
You can read and watch English all day and still freeze when you have to speak it. Speaking needs a partner, and human partners have limits: tutors cost money per hour, and friends are patient exactly once. SpeakEasy is a conversation partner that is always free at 11pm and has never once laughed at anyone’s accent.
How it works
The Expo mobile app joins a LiveKit room over WebRTC, and an AI agent gets dispatched into the same room. From there it is a live voice loop: your speech goes through Deepgram for transcription, GPT-4.1 decides what to say, and Cartesia turns the reply into a voice. The agent reads your profile from the room metadata, so an A2 beginner and a C1 speaker get noticeably different conversation partners.
The part I care most about is what the tutor does not do. It does not interrupt. While you talk, the model silently files grammar errors, pronunciation flags, and vocabulary suggestions through function tool calls. When the session ends, the transcript and all that feedback land in Supabase, and you get a review screen with the corrections. Mid-sentence correction kills fluency; this keeps the conversation moving and saves the red ink for later.
Status
A working prototype from early 2026, quiet since. LiveKit is self-hosted via Docker, the database is Supabase with row-level security, and running it takes three terminals and a dev build, since LiveKit’s native modules rule out Expo Go. The full pipeline story is in the case study.