System planProposal under discussion
A real estate agency on autopilot
AI answers clients on WhatsApp at any hour. The team works from Telegram. The owner follows everything on a dashboard.
AI answers clients on WhatsApp at any hour. The team works from Telegram. The owner follows everything on a dashboard.
imobia is the plan for an in-house client service system for a real estate agency that handles sales and rentals with 5 agents. The client sends a message on WhatsApp and an AI replies right away, understands what they are looking for and shows only properties that exist in the listings. When it is time to visit or negotiate, the conversation moves to the agent, who replies from Telegram. There are no forms to fill in: whatever can be automatic, is. This page records the proposal; the system has not been built yet.
The first recommendation combined EspoCRM, Chatwoot, n8n and Meta's official API. We cut everything that added work without adding results and kept a single in-house app.
The AI replies 24/7 on WhatsApp, asks what matters (buy or rent, area, price range, bedrooms) and sends photos of real properties.
The agent gets the lead with a summary, replies from Telegram and the message reaches the client's WhatsApp. The AI goes quiet in that conversation.
A read-only dashboard shows the funnel, conversations, properties, visits, each agent's numbers and how much the AI is costing.
Everything runs on a VPS with Docker: the Evolution API (the bridge to WhatsApp), the imobia app and a Postgres database.
# proposed architecture
Client (WhatsApp)
│
Evolution API ──► imobia app (Node/TypeScript) ──► Postgres
│ • AI agent with tools
│ • property × client match
│ • follow-up and reminder schedule
▼
Telegram bot (team) Web dashboard (read-only)
Ideas taken from existing projects (imob360, wacrm, EspoCRM, Chatwoot), rewritten our own way.
Finds the client by phone number, joins split messages and transcribes audio on the machine itself.
Uses tools: search properties, send photos, offer a visit, call an agent. It never mentions a property that did not come from the database.
The conversation changes state: AI → waiting for agent → agent → closed. The agent gets a ready summary.
The agent sends photos and a voice note on Telegram. The AI fills in the fields and asks for confirmation with a button.
A new property came in? The system tells the agent about past clients who were looking for something similar.
Visit reminders and follow-ups on stalled leads, only during business hours and stopping at the first “not interested”.
An example of the planned flow. The messages are illustrative.
Nobody on the team is working. The AI replies right away.
# WhatsApp Client: Hi, I want to rent a 2-bedroom apartment downtown, up to 2,500 AI: Hi! I have 3 options downtown in that range. Shall I send you the photos?
It searches the listings, sends the photos and offers visit times. If nothing matches, it says so and offers an agent.
A summary arrives on Telegram. There is no need to ask the client again what they want.
# Telegram — new lead Lead: Ana · rental · downtown · 2 bedrooms · up to R$ 2,500 Interest: AP-0142, AP-0157 · suggested visit: Sat 10am # the agent writes here and the message goes out on Ana's WhatsApp
The agent sends photos and speaks the details. The AI builds the listing and asks for confirmation. Then it notifies whoever was looking for something like it.
On the dashboard: how many leads came in, how many turned into visits, who is stalled and how much the AI cost this month.
This is the choice still needed to close the design. The other parts work with any option.
One agency group with topics. Each lead becomes a topic, the conversation is mirrored there and the agent replies right in it. All 5 see everything and can cover for each other.
Leads are assigned in rotation and each agent sees only their own. To reply, they have to point out which client it is. More private, more confusing.
A group for the owner to follow and a private chat for each agent. More complete, with twice the pieces to maintain and explain.
Points raised during research, with the safeguard planned for each one.
| Risk | Safeguard |
|---|---|
| The Evolution API is not official and the number may be banned | Dedicated SIM, only reply to people who wrote first, human pace, opt-out option, a path to the official API if needed |
| AI making up a property | It only talks about what comes from the database search |
| AI cost growing | Cost logged for every reply and a monthly cap |
| iPhone audio failing on some versions | Ask for text when the audio cannot be read |
| Personal data (LGPD) | Purpose notice, minimal data, local transcription, retention period |
Instagram, website, listing portals, contracts and rent collection are left out of the first version.