AI product · Co-founder
Parivar AI
I built the Node.js backend for a Hindi and Hinglish companion app. LiveKit handles voice calls, and PostgreSQL with pgvector stores conversation memories.
By Krrish Kumar
On Parivar AI, someone can talk to Dadi in Hinglish today and continue the conversation tomorrow. My part was keeping a persona’s voice, conversation history, and memory consistent across text, calls, and photos.
How it works
- Parivar AI appFlutter · Android
- Node.js backendsessions · pgvector memory
- Persona LLMGemini → OpenAI → Groq
- Voice callLiveKit · WebRTC
Capability map of the shipped product. I built the backend; the Flutter client was built by my co-founder.
Why it's built this way
AI behind a service boundary
The client never talks to a model directly. One backend owns prompts, context, and provider calls, so personas stay consistent and providers can change without an app release.
Language as a product requirement
Hindi and Hinglish are the default, not a translation layer. Replies and voice are tuned for how families actually talk.
Honest by design
The product states plainly that users are talking to an AI and positions itself as a companion, never a replacement for real family.
Outcome
Launched on Google Play with 100+ downloads; currently being redeployed. I built the Node.js backend, deployed on AWS EC2 (PM2, Nginx) with self-hosted PostgreSQL and LiveKit; my co-founder built the Flutter client.
How it works, step by step
- 01The user picks a persona and talks to it in Hindi or Hinglish, by text or by voice.
- 02The app sends every interaction to the Node.js backend, which holds the session, the user's context, and long-term memory in PostgreSQL + pgvector.
- 03The backend calls the LLM for an in-character reply, falling back from Gemini to OpenAI to Groq; voice calls run through a LiveKit agent over WebRTC.
- 04Photos, memories, and reminders run through the same service, so every persona behaves consistently across features.