AI Property Search Assistant

The Problem

Property seekers had to dig through listing sites, WhatsApp groups, and rigid filter forms to find what they wanted — and most of Nigeria's real listings live informally in WhatsApp groups, not on any structured platform.

The platform needed a way to let users just describe what they want in plain language, backed by a live inventory that was actually being updated from those same informal sources.

The Approach

This assistant is the solution built out of this problem

The assistant is a chat-first search experience: users describe what they want in plain language, and the system runs a hybrid search — structured SQL filters combined with Voyage AI vector embeddings for semantic matching — rather than relying on keyword filters alone. Budget is handled deliberately: if a user says "under ₦50m," the search expands in one direction only, never silently drifting the ceiling upward. If a location returns too few results, the assistant automatically discovers and searches neighbouring areas before giving up.

A hard rule keeps the assistant honest: it's explicitly forbidden from narrating individual listings from memory. Every property description shown to a user comes from the search tool's own pre-formatted output, never generated free-form — the single biggest source of hallucinated listings in a system like this, closed off by design. A guardrail enforced in code beats an instruction you hope the model follows. When a search is genuinely exhausted, the assistant asks the user before escalating to a human agent, and hands off the full conversation context rather than starting the agent from zero. Off-topic messages get redirected gracefully instead of breaking the flow or stonewalling the user.

On the supply side, listings come from a mix of formal and informal sources such as WhatsApp groups, with no proper structure. I designed and built an automated ingestion pipeline that pulls from these sources continuously, feeding the database roughly 1,000 new listings a day.

• Solo-built the full stack and AI component
• Automatic property ingestion pipeline from formal and informal sources
• Hybrid search — structured SQL filters + Voyage AI vector embeddings for semantic intent matching
• Anti-hallucination guardrail: every listing shown to a user comes from the search tool
• Automatic neighbourhood expansion when a search returns too few results
• Escalates to a human agent with full conversation context

The Outcome

(Images shown with the client's consent.)

besides the mobile view, the product is also accessible via desktop