The challenge
Traditional real estate search is fundamentally broken for today's Indian homebuyers. Existing portals force buyers into rigid dropdown filters, bedrooms, budget in lakhs, PIN code, builder, when real intent is nuanced: "a quiet 2-BHK with morning sun in a family-friendly neighbourhood near a good school, metro access, ready-to-move-in, under ₹1.5 crore." Listing photos go unanalyzed. Most Indian homebuyers think in Hindi, Telugu, Tamil, or Kannada, but every portal forces English. Geography gets treated as a bounding box on a map, not as commute, school, market, and metro context.
What we built
The platform end-to-end on AWS, an India-first real estate search experience that understands what homebuyers actually want, in the language they actually speak. Buyers describe what they want in plain language; Amazon Bedrock foundation models handle intent extraction and ranking, with multimodal models (Claude with vision, Nova) analyzing listing photos, text descriptions, location data, and amenities together, the AI understands "a bright kitchen" by looking at the actual photos. Sarvam AI, a foundation model purpose-built for Indian languages, is integrated directly into the natural-language search flow so a buyer asking in Hindi, Telugu, Tamil, or Kannada gets a response in the same language. Retrieval runs over Amazon Aurora PostgreSQL with pgvector, a single managed database for listing data and embeddings. Personalization, conversational refinement, and geospatial reasoning round out the experience.
Why this matters
The Sarvam AI integration means homebuyers in Hindi-speaking, Telugu-speaking, Tamil-speaking, and Kannada-speaking households can search in their first language, a meaningful access lever in a market where every existing portal forces English-only interaction. The multimodal stack on Amazon Bedrock turns listing photos into a queryable layer rather than passive scroll material. Aurora PostgreSQL pgvector retrieval keeps the operational stack to a single managed database. Buyers stop scrolling through hundreds of nearly-identical listings and start having a conversation with a search engine that understands them.