English Circle

An AI English-conversation tutor for Mexican Spanish speakers. Live group voice classes with “Alex,” an AI teacher who understands Mexican slang, corrects you gently in real time, and sends a personal report after every class.

Don’t have a code? Request a code →

The problem

Millions of Mexicans want to speak English, but the options fail at the one thing that matters — real conversation. Apps gamify vocabulary but never make you talk; private tutors cost a fortune; and generic chatbots have no structure and no cultural awareness (they flag Mexican slang as “errors”).

What I built

A conversational teacher that runs on a 25-agent pipeline (Google ADK + Gemini 2.5 Flash) and a dual-path voice architecture for natural, low-latency speech:

• Fast path — Gemini Live handles the real-time spoken conversation (~500 ms).
• Background path — the full pipeline runs asynchronously, feeding gentle corrections, topic navigation, and safety checks back into the live conversation.
• Mexican-Spanish aware — a localism registry so slang like chido or no manches is welcomed and bridged to English, never flagged.
• Per-session reports — every class ends with a personal report: corrections, expressions used, and progress.
• Group classes — LiveKit voice rooms with breakout pairs, so everyone actually speaks.

Engineering highlights

• Multi-agent orchestration (interpretation, safety routing, error tracking, multi-critic quality loop) built as composable Sequential / Parallel agents.
• A cost/latency optimization: a trimmed “demo mode” pipeline that cut per-turn LLM calls ~80% after profiling production logs.
• Real-time transport on LiveKit; deployed as a one-image, two-role (API + worker) service on Railway, with a Next.js front end on Vercel.
• 600+ tests across the agent pipeline and voice layer.

Stack