Smart Shaadi
Problem
Matrimonial platforms match on filters, not on how two people actually communicate or whether their families align.
My contribution
I built the platform end to end — profiles, authentication, chat and payment integrations — and designed and wired its AI layer.
Implementation
- Streaming, multi-turn assistant with conversational memory and read-only tool calling.
- Provider-agnostic AI layer supporting Gemini and Claude.
- Semantic profile re-ranking with embeddings in PostgreSQL / pgvector.
- Conversation coaching, emotional scoring and experimental compatibility indicators.
Technical details
Semantic profile re-ranking uses 768-dimensional embeddings. Six user-facing AI features are wired end to end: a conversation coach, a sentiment-based emotional score, an experimental compatibility-risk indicator, family-values compatibility, the streaming assistant and semantic re-ranking. Further backend/admin models for churn risk, response-rate reputation, marriage-readiness and profile optimisation are implemented without a public UI.
Current status
Pre-launch: company registration is pending and the environment currently runs on dummy data and services ahead of commercial launch. The compatibility and risk indicators are experimental, not scientifically validated.









