Agent²

Your real estate agent AGENT

Comment 1

Inspiration

WATCH FUNNY SKIT: https://youtu.be/vdsDCTusscE

House hunting today means opening Zillow, applying filters, and scrolling forever.

Real estate agents don’t help much unless you’re about to make them money.

So we built an agent anyone can use.

Call a number, describe your dream home, and our system sends you matching listings by text.

Finding housing online is still surprisingly inefficient. Most platforms require users to open multiple websites, apply filters, browse dozens of listings, and manually compare results. Even with modern tools, the process still feels like filling out forms rather than simply describing what you want.

At the same time, large language models have become very good at understanding natural conversation. We started wondering what the home search experience would look like if it worked more like talking to a real estate agent.

Instead of forcing users to navigate websites and filters, we wanted people to simply call a number and describe the home they are looking for in their own words. The system would listen, understand their preferences, and return matching listings directly by text message.

Agent² explores this idea by combining conversational AI, structured data extraction, and real estate listing aggregation into a single experience that feels as simple as talking to someone.

What it does

Agent² is an AI powered real estate assistant that lets users search for homes through a phone call.

A user calls the number and describes what they are looking for, such as budget, number of bedrooms, preferred neighbourhood, or deal breakers. The system processes the conversation, extracts structured search criteria, and finds relevant listings.

The results are then sent directly to the user's phone by text message.

This removes the need for apps, accounts, or search filters. Instead of navigating a real estate website, users simply describe the home they want and receive matching listings within seconds.

How we built it

Our app is mainly backend focused, with most of the logic dedicated to processing conversations, extracting search criteria, gathering listing data, and returning results to users.

Backend

We use Telnyx for SMS messaging, enabling back-and-forth communication with users and scheduling calls between users and our AI agents. Once a call is scheduled, users speak with our PersonaPlex voice agent, which processes speech and forwards conversation data to the backend using FastAPI. PersonaPlex is deployed on AWS.

After the call, we leverage multiple Railtracks endpoints powered by LLMs to extract housing preferences from the conversation and generate ranked listings tailored to the user. Listing data is scraped from real estate sites using Playwright and ScraperAPI for browser automation and proxy rotation, then parsed and structured with BeautifulSoup and Pandas.

Frontend

Our landing page (the entry point for connecting users to the service) is built with plain HTML, CSS, and JavaScript. The site is deployed statically on Vercel with a .tech domain.

Challenges we ran into

Real estate data access
One of our biggest challenges was obtaining reliable real estate data. We experimented with several real estate APIs, but quickly ran into limitations:

  • Many of them only returned sample data for free accounts, while the paid plans that provided real listings were extremely expensive and not practical for a hackathon.
  • Several APIs focused primarily on the United States and returned incomplete or irrelevant results for Canadian listings.

Because of these limitations, we had to rely on live scraped data instead. This created additional challenges since listing data across websites is often inconsistent and structured differently, requiring careful parsing and normalization.

This difficulty in accessing reliable housing data is exactly part of the problem Agent² aims to solve.

Voice agent accuracy in loud places
Audio from louder environments could confuse the voice agent and lead to incorrect or incomplete preference extraction.

Latency across the pipeline
Another challenge was latency. Because the system depends on multiple steps, keeping the overall experience responsive was important. We had to think carefully about how to make the interaction feel smooth enough to still feel conversational.

Accomplishments that we're proud of

  • A full call to text workflow
    Users can call, describe the home they want, and receive matching listings by text without browsing websites themselves.

  • Making AI useful in a practical way
    Instead of using AI just for the sake of it, we built a system where it solves a real problem by turning natural conversation into structured real estate search criteria.

  • Working around real world data limitations
    When APIs turned out to be too limited, too expensive, or not relevant enough for Canada, we adapted and still built a working solution that could return useful listings.

  • End to end integration in a hackathon setting
    We connected Telnyx, FastAPI, railtracks, LLM extraction, live listing collection, and SMS delivery into one working product within a very short time.

  • The project itself :)

  • ...And our demo video

What we learned

  • Structured extraction matters a lot
    People do not describe homes in a neat form, so turning natural conversation into something reliable enough to search with is harder than it first sounds.

  • Real estate data is harder to access than we expected
    Data quality, cost, and regional coverage can completely shape a product, especially in Canada.

  • Voice interfaces are powerful, but they come with tradeoffs
    Background noise, unclear phrasing, and latency all affect the experience much more than in a normal text based app.

  • User experience matters just as much as the backend
    Even if the system is doing a lot of work internally, the interaction has to feel simple and fast enough for users to trust it.

  • Parallel team work was important
    This project only worked because scraping, backend, AI, and product work could all move forward at the same time.

What's next for Agent²

There are many directions we would like to explore next.

We would like to improve the listing aggregation system to include more data sources and richer property information.

We are also interested in adding preference learning so the system can improve its recommendations over time as it learns what types of listings users interact with.

In the long term, Agent² could evolve into a fully automated real estate discovery assistant that handles everything from initial search to scheduling viewings.

Built With

Share this project:

Updates