HouseBuyerAgent

Inspiration

My personal experience looking for a home to buy in this correct environment made me spend too much of my time researching a lot about the various factors that went into deciding which home to buy and I decided to build something that does the work for me and others

What it does

This application provides an intelligent home buying assistant that coordinates 5 specialized AI agents to help users find, analyze, and get personalized property recommendations. The system processes real estate data, performs semantic search, and provides comprehensive property analysis through a modern React web interface.

How we built it

Built with ADK, Leveraging Github Copilot Agent mode. Set out to create a multi agent home buying platform that can help do the searching, research and recommendation. For the purpose of the hackathon MVP, generated Synthetic datasets and uploaded to Google Big Query. Created Embeddings for the listings to allow for Semantic search in addition to keyword search. Data generated from the listings agent is sent to locality agent, hazard agent and affordability agent parallelly to generate scores for each top listing. the recommendation agent generates a final recommendation after reviewing the scores received from the 3 agents for the top 5 listings. agents leverage google Gemini flash 2.0 through vertex AI. Query history is stored in a Firestore database for easy analysis . The flask api and the react frontend are deployed to Google Cloud Run

Technology Stack Summary

Challenges we ran into

Multiple challenges were encountered. some of them are listed below. Vector Embedding support on Google BigQuery. it only supports 768 dimensions so had to pick a specific embedding model or reduce dimension to compare. ADK framework being very new is not well known to LLM and Needs websearch for models to be able to troubleshoot and generate code correctly. Deployment to Cloud run directly from Github caused issues with the cloud build triggers which were eventually fixed by deploying through Google cloud cli.

Accomplishments that we're proud of

Building Multi Agent system with parallel fan-out /gather and sequential design pattern for home buying experience. building a react frontend to capture user queries and store them on a highly scalable and performant firestore db for future retrieval and analysis. implementing vector search with bigquery which is very new and has some limitations tested the limits of the implementation.

What we learned

Google Cloud Run, Cloud Build, Big Query, Firestore, Vertex AI, IAM and roles, ADK This project gave me a lot of opportunity to develop using Google cloud and google AI tools. My expertise is mainly on Azure and AWS.

What's next for HouseBuyerAgent

The MVP is built with Synthetic data. I want to integrate with realtime API for listings , Fema Hazard analysis, School ratings and other data to build a live production website that can help first time home buyers take the guess work and stress out of the home buying experience

Built With

• adk
• bigquery
• cloudrun
• firestore
• flask
• google-cloud
• python
• vertex