Problem

Learning to speak a language needs live conversation and immediate feedback on how you actually sound — not flashcards. Parker AI is a conversational Japanese tutor built to hold that spoken practice and correct it as it happens.

Approach

The core is a real-time voice pipeline: speech recognition (Deepgram Nova-3) feeding straight into LLM response generation (Claude), so a spoken turn becomes a spoken reply. That chain is production LLM engineering rather than a demo — the work sits in the latency budgets, evaluation sets and cost control that keep a two-model loop responsive and affordable at scale. It’s the seventh platform in a sequential build record, the one chosen to add production AI to the stack, running on Next.js 15 and TypeScript with Supabase behind it and deployed on Vercel.

What it does

Real-time speech recognition into conversational LLM responses
Pronunciation scoring on spoken input
A JLPT-aligned curriculum
Spaced repetition to schedule review

Status

In development. The focus is the production LLM engineering underneath the tutor — the speech-to-LLM latency budget, the evaluation sets that hold response quality, and the cost controls on the pipeline.