Larlink: AI agents for real estate
WhatsApp lead qualification against inventory, with orchestration flows and a human handoff dashboard.
Overview
Larlink helps real estate agencies qualify leads on WhatsApp. The bot answers early questions, searches inventory with semantic matching, and hands a structured profile to a human agent. I worked on the Windmill orchestration and built a React dashboard so agents can watch chats and take over when a lead is hot.
Problem
Agents spend hours on the same early questions about budget, location, and timing. That slows response and leaves less time for closing.
Constraints
- Everything had to stay on WhatsApp. No app install for the buyer.
- Replies had to read as chat, not a phone tree.
- Agents needed a way to monitor the bot and take over instantly.
Architecture
Windmill for orchestration, Gemini for language, Supabase pgvector for property matching. I focused on the conversation flows and the handoff UI: a React dashboard that looks like WhatsApp, talks to Windmill, and lets an agent pause the bot and reply in the same thread.
Key Decisions
Windmill as the orchestrator
Central place for flows, schedules, and API endpoints. When an agent replies in the dashboard, Windmill calls the Meta WhatsApp API. More flexible than the n8n setup we tried first.
Gemini with tool calling
Instead of a fixed script, the model chooses when to reply and when to call search_property, with no phone-tree menu.
Two-stage property matching (pgvector + AI filter)
Vector search finds conceptually close listings (for example quiet place ≈ low noise). A second pass with the model keeps only the strongest matches.
Custom handoff dashboard
Sales still need a human. The dashboard lists active chats. Reply pauses the bot and sends the agent message to WhatsApp.
Tech stack
- Windmill
- Gemini AI
- Supabase (pgvector)
- TypeScript
- WhatsApp API (Meta)
- React / Next.js
Impact
The bot can qualify and match inventory outside agent hours. When an agent takes over, budget and preferences are already summarized.
Learnings
- We tried n8n first. Windmill's code-first model fit complex AI state better.
- pgvector plus a final AI filter beat keyword search alone for property matches.
- In B2B, the handoff UI matters as much as the bot. A WhatsApp-like screen made testing with agents easier.
Live product
Live at larlink.com.br. Larlink marketing site: AI WhatsApp intake for real estate agencies.