A WhatsApp property assistant running on a real agent graph
A WhatsApp assistant for a Hyderabad property platform that helps buyers search, connects sellers to agents, handles loan enquiries, captures listings and manages wishlists — built as a stateful multi-flow agent rather than a chatbot.
Client: Gazam Associates · Hyderabad, Telangana · Status: live
The challenge
Property enquiries arrive on WhatsApp at all hours and each one needs a different journey: a buyer wants to filter by locality and budget, a seller wants to list, someone else wants a loan or an agent introduction. A single linear chatbot cannot hold those apart, and losing the thread mid-conversation loses the lead.
What we built
We built the assistant on LangGraph with Redis-backed checkpointing, so every conversation has durable state and can be resumed exactly where it stopped. A supervisor routes each message into the right flow — buyer search, find an agent, loan, post a property, agent registration, wishlist, your listings, support — and each flow owns its own steps while sharing one warm, consistent persona across all of them.
Engineering highlights
- Supervisor-routed multi-flow architecture: eight distinct journeys behind one WhatsApp number.
- Redis checkpointing so a conversation survives restarts and a customer can pick up hours later.
- Production hardening after a real incident — Redis is given an explicit memory ceiling and eviction policy on every boot, and orphaned checkpoint index keys are swept, because an unbounded instance was once OOM-killed.
- Formatting written specifically for WhatsApp: bold labels, numbered options, short paragraphs and no walls of text.
- The assistant identifies itself as an AI assistant and never claims to be human.
- Covers the full Hyderabad market by locality — Gachibowli, Hitech City, Kukatpally, Madhapur, Banjara Hills, Narsingi and more.
Technology
Services used
Want something like this?
Tell us your problem and we'll tell you exactly how AI solves it and what it costs.