VoiceMirror

● Overview

Duolingo-style lessons for how you talk: a skill-tree learning path, AI-scored practice, and free practice with a text coach, a voice coach and AI avatar calls. Native iOS with one backend behind it.

● The hard part

Coaching only works if feedback is fast, safe and personal, and a mobile app can't hold provider keys. Every model, transcription and speech call had to move server-side without making the app feel slower.

iOS

native SwiftUI

19

node skill tree

0

provider keys on device

● What I built

01

SwiftUI app

An iOS 17+ app with a skill-tree Learn path, XP, levels and streaks, voice practice and avatar calls, plus RevenueCat subscriptions and PostHog analytics.

02

One AI gateway

A NestJS API owns auth, content, scoring, caps and the paywall, and is the only thing that talks to OpenAI and ElevenLabs. The app ships zero provider keys.

03

Keyless development

A deterministic fake AI provider runs every route with no external keys, and config refuses it outside development, so the whole product can be built and end-to-end tested offline.

04

From microservices to one API

The first version was a Go middleman with a safety gate in front of Python text, voice and avatar coach services with RAG memory. I collapsed it into a single backend that's easier to ship and to reason about.