top of page

Our Methodology

Our Motivation

Buying a home is not only one of the largest financial decisions a person will make in their life but can also be one of the most overwhelming and stressful. New real estate laws starting in 2025 will only increase the financial burden placed on many home buyers by requiring buyers to pay their agents directly. Our project aims to empower users to take control of their homebuying experience by providing them with expert level guidance and support at their fingertips without all the hidden fees.

Data

To power Brooke and her tools, we made use of three key datasets: 

  1. Training Data: Zillow’s real estate queries

  2. Housing Data: Redfin’s housing market data

  3. Financial Data: Tax brackets and tax rates from taxfoundation.org, mortgage rates from Freddie Mac MBS, and homeowners insurance costs from Policy Genius.
     

The data from Zillow consists of over 20,000 synthetically generated questions and legally compliant responses covering a range of real estate-specific topics. The Redfin data consists of historical home listing prices, home sale prices, active listings, number of days on the market, and more relevant KPIs. The data from Freddie Mac and Policy Genius includes average insurance costs by Zip, State + Federal Tax Brackets, and up-to-date mortgage rates by State + credit score.

​

All of our data was quality-checked for nulls, most recent date (2024+), and accuracy. We conducted EDAs to ensure that values were as expected for data regarding mortgages, insurance rates, and housing prices. We also confirmed that the training data from Zillow was comprehensive for the 108 distinct topics covered, and no clear topics were missing or lacked adequate coverage.

Approach

Brooke is designed to provide a holistic view of the real estate market while providing personal support and relieving the stress that comes with buying a home through three key features : 1) Chat with Brooke 2) Budget Calculator 3) Market Trends. 
 

Chat with Brooke

To power conversation with Brooke, we opted to use the open source model DeepSeek-R1 Distill LLaMa 8B because of its reasoning skills as well as its fast and efficient processing. This ensures high-quality responses without the wait time associated with a live broker.
 

To ensure our chatbot generates both helpful as well as legally compliant responses aligned with the Fair Housing Act and the Equal Credit Opportunity Act, we fine-tuned the model using three key datasets: 

  1. Zillow’s dataset of real estate-focused queries and responses generated using GPT-40 

  2. Our proprietary dataset of compliance-safe responses generated using Claude (Anthropic)

  3. Redfin property data (integrated through a RAG pipeline, along with the specific user inputs through an introductory questionnaire)

To optimize performance, we fine-tuned the model using Unsloth which allows for efficient memory, using 70% less memory and running 2x faster while maintaining reliable model performance.

​

Budget Calculator

For our Budget Calculator, we loaded and cleaned regional-level insurance and tax data, and aggregated housing cost data by region to calculate appropriate ranges. We also established a classification process based on property type to interact with user inputs in the initial questionnaire. 
 

The tool is powered through Gradio, with expert review by both a real estate agent and financial expert separately.
 

Market Trends

For Market trends, we loaded and cleaned city, state, and metro-level KPIs, and created additional aggregations for 3 and 6-month rolling averages to power a more readable visualization. Similarly to the Budget Calculator, the Market Trends product is default populated based on the user inputs in the initial questionnaire to provide a more seamless user journey. 

The visualization is powered through Gradio and Altair.

Evaluation

To ensure that chatting with Brooke delivers accurate and compliant results, we conducted an evaluation process comparing Brooke to other open-source models such as:

  • LLaMA3

  • Mistral Small

  • DeepSeek-R1 (8B Distilled)

We benchmarked all of the models against Claude 3 Sonnet. We opted to use Claude 3 Sonnet because of its reasoning abilities while also producing safe and compliant answers at a high speed. We tested against 2,000 real estate-focused queries, half of which were focused on safety/compliance and half that prioritized the usefulness of the model. We chose to measure performance by using the BLEURT Score because it is excellent at focusing on the content of the responses rather than prioritizing exact word matches as well as its effectiveness for evaluating longer conversations with Brooke. 

Our model outperformed all 3 models that we compared it to across both safety and usefulness indicating that Brooke provides both highly accurate and legally compliant responses

bottom of page