Mobile app

AMOR

Talk to an AI companion that actually listens — tap once and have a hands-free conversation, no push-to-talk.

Platform: iOS + Android (React Native, Expo)
Role: Full-stack mobile developer, solo
Type: Client project, delivered in milestones
Timeline: Apr – Jul 2026 (~3.5 months)
Status: Feature-complete · prepared for store submission

// overview

AMOR is a mobile AI companion app for iOS and Android. The idea is simple: instead of typing at a chatbot, you can call it — tap once and have a natural, hands-free conversation, the way you would talk to a friend on the phone.

I built the product end to end: the React Native front end, the Supabase backend and database schema, the server-side voice pipeline, the authentication and guest system, and the Stripe payment and wallet layer.

The hard part of a voice app is not the AI — it is making the conversation feel natural. Most voice apps make you hold a button while you speak. AMOR does not. A custom Voice Activity Detection system continuously reads the microphone level and works out on its own when you have started speaking and when you have finished. It waits out short pauses, rejects background noise, and ends the recording precisely when you stop, so the loop is completely hands-free: you speak, it detects silence, it thinks, it replies, it listens again.

Users can try the app instantly as a guest with small daily limits, then create an account — at which point all their guest conversations and call history are migrated into the new account by a single database function, so nothing is lost. Registered users get a wallet: chat messages cost a cent each, voice calls bill at $1.00/hour charged per second actually used, topped up with real Stripe payments.

Scale

7

AI companion personalities

20+

Tables under RLS

2

Edge functions

~40

TypeScript source files

8

Milestones delivered

100 ms

VAD sampling interval

// what it does

Key features

Conversation

Text chat and hands-free voice calling with the same AI companion
Seamless switching between chat mode and call mode mid-conversation
A guided "first call" onboarding script that introduces the companion and learns the user’s name
Full call history with replayable transcripts

Voice engine

Custom Voice Activity Detection — no push-to-talk; speech start and end are detected from live audio metering
Tunable dB threshold, silence duration and minimum-speech guards
Server-side pipeline: Whisper speech-to-text, GPT-4o-mini, OpenAI TTS
Audio playback with a per-companion voice

AI personalities

Seven distinct companions — Luna, Aria, Alex, Marcus, Sienna, Theo and Iris
Each with its own system prompt, personality profile, TTS voice and avatar
An accent colour per companion that re-themes the UI
Per-companion memory: each builds its own long-term context about the user

Accounts & data

Guest mode with device-based sessions — usable in seconds, no signup wall
Email and password auth via Supabase, with real-time password-strength validation
Forgot-password and reset flows over an amor:// deep link
Automatic guest to registered-user data migration, and Row Level Security on every table

Payments

Stripe wallet with $4.99 / $9.99 / $19.99 top-up packages
Per-second billing for voice calls, per-message billing for chat
Server-verified, idempotent crediting — the Stripe secret key never leaves the backend
Pricing driven from a pricing_config table, so rates change without an app update

Polish

Splash, onboarding carousel, auth gate and main app flow
Three-tab layout (Chat · Calls · Account) with badges and active states
Offline-tolerant local persistence via AsyncStorage
Branded transactional emails, plus Privacy Policy, Terms and Help screens

// engineering

The hard parts

The problems that shaped the architecture, and how each one was solved.

01

Making voice feel like a real call, not a walkie-talkie

Fixed-duration recording either cut people off mid-sentence or left awkward dead air. The VAD service samples the microphone’s metering level every 100 ms, requires roughly half a second of sustained speech before it counts as "the user is talking", then ends the turn after 1.5 s of silence. The result is a fully hands-free conversation loop with thresholds that can be tuned for noisy environments.

02

Running a paid AI pipeline from an app that cannot hold secrets

Every AI and payment call moved into Supabase Edge Functions. The client sends audio and gets audio back; the OpenAI and Stripe keys exist only as server secrets. This also made pricing changes and model swaps possible without shipping a new app build.

03

Letting people try the app without losing their history

Guests get a device-scoped session and small daily limits. The moment they sign up, a single migrate_guest_to_user() database function re-parents every conversation, message and call record to the new user ID — designed to be safe to run more than once.

04

Billing accurately for something measured in seconds

Wallet balance and every deduction live in Postgres behind SECURITY DEFINER RPCs, so the client can never write its own balance. Calls bill per second at an hourly rate read from a config table, and top-ups are credited only after the PaymentIntent is re-fetched from Stripe and verified for status, amount and owner — keyed by payment ID, so replays are harmless.

// stack