AI Solutions Architect

How I Built an AI Agent for Monitoring the Real Estate Market in Georgia

A case at the intersection of RAG, Low-code, Python, and Go.

Python (Telethon, Playwright) Go (Gin Premium) Gemini 1.5 Flash Docker Compose Streamlit

01. Introduction

The real estate market in Georgia (especially in Batumi and Kobuleti) is a chaos of hundreds of Telegram channels and Facebook groups. The information there is unstructured, noisy with ads, and quickly becomes outdated.

I set a goal: to build an MVP intelligent agent that collects data from two platforms in real time, analyzes it with an LLM, and provides accurate insights without “hallucinations.” In this project, I acted as an AI Solutions Architect, using the Antigravity agent to accelerate infrastructure assembly, which allowed me to focus on RAG pipeline logic.

02. System Architecture: Multiplatform RAG

INGESTION LAYER
AI PROCESSING LAYER
STORAGE & RETRIEVAL
USER INTERFACE
Telegram (Telethon)
Facebook (Playwright)
Gemini 1.5 Flash
Intelligent Parsing & Entity Extraction
SQL Database
Structured Metadata
RAG Engine
Semantic Filtering
Streamlit Dashboard
Interactive Analytics

The heart of the project is a Dynamic RAG (Retrieval-Augmented Generation) architecture. Unlike typical chatbots, my agent does not generate answers from thin air. It builds responses based on fresh “raw” data from social networks.

03. The RAG Pipeline

1. Data Ingestion

The system connects to Telegram sockets and scrapes public FB groups in Kobuleti and Batumi in real-time.

2. Intelligent Parsing

Each message is sent to Gemini. The AI extracts entities: city, price, rooms, and author type (owner or realtor).

3. Storage & Metadata

Data is stored in an SQL database with full metadata, enabling hybrid search (semantic + exact).

4. Augmented Generation

When users ask specific questions, the system fetches matching records and injects them into the prompt for fact-based answers.

05. Engineering Challenges

Self-Healing and Fault Tolerance

Developing in regional conditions (Batumi/Kobuleti) required accounting for risks of power and internet outages. To ensure the MVP does not require manual restarts, I implemented:

  • Docker Restart Policies Automatic container recovery after critical failures ensures 24/7 monitoring.
  • Network Resilience Python services include exponential backoff reconnection logic. If the internet drops, the system waits and seamlessly resumes.

06. Results & Expertise

80%
Spam Filtered
3
Cities Covered
100%
Fact-Verified
MVP
In A Few Days

This project demonstrates how the combination of AI tools (Antigravity) and solid architecture enables a single engineer to build an enterprise-grade product in just a few days.

"The future of development is not in writing lines of code, but in orchestrating intelligent data flows and AI architecture."

Interested in similar AI solutions?