Case Study

Building Nona

AI Real Estate Concierge

4+ Months
Development Time
124K+
Lines of Code
8
AI Agents
38K+
Users Reached
Project Overview

The Challenge

Finding a home to rent or buy in the Philippines is time-consuming, fragmented, and risky. Listings are often duplicated, incomplete, or outright scams. Home seekers waste hours browsing unverified sites and contacting multiple brokers without a clear way to compare or validate options.

Traditional real estate platforms in the Philippines suffer from several critical issues:

  • Fragmented Information: Property listings are scattered across multiple platforms, making comprehensive searches nearly impossible.
  • Trust Issues: Scam listings and fraudulent brokers are rampant, eroding user confidence.
  • Time-Consuming Process: Home seekers spend days or even weeks sifting through irrelevant listings.
  • Poor User Experience: Complex filter systems and unintuitive interfaces frustrate users.
  • Limited Context: Existing platforms rarely provide local insights about neighborhoods, commute times, or nearby amenities.

The Solution

We built Nona, an AI-powered real estate concierge that acts as a trusted assistant for home seekers. Instead of clicking through dozens of buttons and listings, users chat directly with Nona to get personalized, verified property recommendations.

Multi-Agent Architecture

Built on Google's Agent Development Kit (ADK), Nona orchestrates 8 specialized AI agents to handle complex real estate conversations with precision. Each agent is purpose-built for specific tasks:

Core Capabilities:

  • Natural language property search & filtering - Users can describe their ideal property in plain language
  • Multi-agent intent routing & conversation management - Intelligent routing to specialized agents based on user intent
  • Intelligent lead capture with context preservation - Automated lead collection at optimal conversation points
  • Real-time property database with 1,200+ listings - Verified property database with comprehensive filtering
  • Web search fallback via Tavily integration - Expands search beyond database when needed
  • Google OAuth & session persistence - Seamless authentication and conversation continuity

Key Impact

  • 16,263 conversations initiated by users seeking properties
  • 276% conversion from visit to active search (4,788 / 6,263)
  • 399% new user rate demonstrating market reach (37,737 / 38,233)
  • 4Zero major outages during 3-month production period
  • 5Sub-second response times for AI agent interactions
  • 6Scalable architecture supporting 44K+ sessions
Technical Approach

How We Built It

Multi-Agent Architecture

Built on Google's Agent Development Kit (ADK), Nona orchestrates 8 specialized AI agents to handle complex real estate conversations with precision.

The 8 Specialized Agents

1. Intent Analyzer - Classifies user messages and extracts entities for intelligent routing to the appropriate specialized agent.

2. Property Search V2 - Orchestrates 5 sub-agents for comprehensive property discovery:

  • Param Extractor - Extracts search parameters from natural language
  • Database Search - Queries verified property listings
  • Web Search - Searches external sources via Tavily
  • Acknowledgement - Confirms search understanding
  • Result Follow-up - Manages result presentation and refinement

3. Property Advisor - Provides expert guidance on real estate decisions and market insights, leveraging RLC knowledge base.

4. Service Agent - Handles service requests (cleaning, repairs, moving) and connects users with verified service providers.

5. Lead Collection - Captures and validates contact information with conversation context for follow-up.

6. Lead Observer - Monitors conversations for optimal lead capture timing without disrupting user experience.

7. General Inquiry - Handles broad questions about real estate, the platform, and general property information.

8. Listing Inquiry - Provides detailed information about specific property listings, including amenities and viewing schedules.

Technology Stack

Nona leverages cutting-edge technologies across the full stack:

  • Google Gemini 2.0 for advanced natural language understanding
  • Google Agent Development Kit (ADK) for multi-agent orchestration
  • FastAPI backend with Python 3.12+
  • Next.js 15 frontend with React and TypeScript
  • Supabase for PostgreSQL database and authentication
  • Tavily API for web search capabilities
  • Sanity CMS for content management
  • Google Cloud Run for scalable deployment
Challenges & Solutions

Overcoming Obstacles

Development Journey: Three Iterations to Production

From proof-of-concept to a fully-featured AI platform in 4 months.

Phase 1: Proof of Concept (August 2025)

Initial fork of SuperAgent framework. Built basic chat interface with mock data, focusing on Manila Metro areas. Established core UX patterns and brand identity.

Key Features:

  • Mock Data System
  • Basic Chat Interface
  • UX Foundation
  • Brand Identity

Phase 2: Version 1.0 - API Integration (September 2025)

Real API integration with live property data. Implemented comprehensive analytics, admin dashboard, and brand refresh. Added Meta Pixel and GTM tracking.

Key Features:

  • Live API Integration
  • Analytics System
  • Admin Dashboard
  • Brand Refresh

Phase 3: Version 2.0+ - Multi-Agent AI (October - December 2025)

Complete multi-agent architecture with Google ADK. Launched Sanity CMS, property feed, Google OAuth, and infinite scroll. 38,000+ lines added in October alone.

Key Features:

  • 8 AI Agents Architecture
  • Sanity CMS Launch
  • Google OAuth
  • Infinite Scroll Property Feed
  • Lead Generation System
AI Engineering

System Architecture

Loading diagram...
AI Agents

Multi-Agent System

🎯

Intent Analyzer

Classifies user messages and extracts entities for intelligent routing

🏠

Property Search V2

Orchestrates 5 sub-agents for comprehensive property discovery

💡

Property Advisor

Provides expert guidance on real estate decisions and market insights

🔧

Service Agent

Handles service requests and connects users with providers

📞

Lead Collection

Captures and validates contact information with conversation context

👁️

Lead Observer

Monitors conversations for optimal lead capture timing

💬

General Inquiry

Handles broad questions about real estate and the platform

📋

Listing Inquiry

Provides detailed information about specific property listings

Technical Scope

Full-Stack AI Platform

⚛️

Frontend

131 React Components

Next.js 15ReactTypeScriptTailwind CSSRadix UIFramer Motion
🐍

Backend

125 Python Modules

FastAPIPython 3.12Google ADKPostgreSQLRedisJWT Auth
🤖

AI/ML

8 Specialized Agents

Google Gemini 2.0Google ADKTavily APILangSmithPrompt Engineering
☁️

Cloud

Multi-Environment

Google Cloud RunDockerSupabaseVercelCI/CD
📝

CMS

32 Content Schemas

Sanity v4.15GROQReal-time CollabWebhooksCDN
🔐

Security

Enterprise-Grade

Google OAuthJWT TokensRate LimitingRLS Policiesbcrypt
Development Journey

From Concept to Production

August 2025

Phase 1: Proof of Concept

Initial fork of SuperAgent framework. Built basic chat interface with mock data, focusing on Manila Metro areas. Established core UX patterns and brand identity.

Mock Data SystemBasic Chat InterfaceUX FoundationBrand Identity
September 2025

Version 1.0 - API Integration

Real API integration with live property data. Implemented comprehensive analytics, admin dashboard, and brand refresh. Added Meta Pixel and GTM tracking.

Live API IntegrationAnalytics SystemAdmin DashboardBrand Refresh
October - December 2025

Version 2.0+ - Multi-Agent AI

Complete multi-agent architecture with Google ADK. Launched Sanity CMS, property feed, Google OAuth, and infinite scroll. 38,000+ lines added in October alone.

8 AI AgentsSanity CMSGoogle OAuthProperty FeedLead System

Impact & Results

6,263
Total Chats
4,788
Searches Submitted
38,233
Unique Visits
37,737
New Users
44,468
Total Sessions

Impact & Usage Metrics

Measurement Period: September 9 – December 9, 2025

Real user engagement data from the first 3 months of production deployment:

  • 6,263 Total Chats - Conversations initiated by users seeking properties
  • 4,788 Searches Submitted - Active property searches
  • 38,233 Unique Visits - Individual users reached
  • 37,737 New Users - First-time platform visitors
  • 44,468 Total Sessions - Overall user engagement

Business Impact

  • 76% conversion from visit to active search (4,788 / 6,263)
  • 99% new user rate demonstrating market reach (37,737 / 38,233)
  • 1.16 sessions per user showing return engagement

Technical Excellence

  • Zero major outages during 3-month production period
  • Sub-second response times for AI agent interactions
  • Scalable architecture supporting 44K+ sessions
  • Comprehensive monitoring via LangSmith and analytics

Platform Capabilities

  • 1,200+ verified listings in property database
  • 8 specialized agents for different use cases
  • Real-time search across database and web sources
  • Intelligent lead capture with context preservation

FrontierAI delivered a production-ready AI platform that exceeded our expectations. The multi-agent architecture handles complex real estate conversations with remarkable precision, and the system scaled effortlessly to support thousands of users from day one.

N

Noneaway Team

Client

Engineering at Scale

58
Git Commits
818+
Files Changed
124K+
Lines Added
55
DB Migrations
214+
Documentation Files
131
React Components
32
CMS Schemas
4
Packages
FrontierAI - AI Solutions for Ambitious Companies