HomeScore

Inspiration

Buying a home is one of the biggest financial decisions a person makes, yet most real estate platforms encourage buyers to filter by price and bedrooms first. This often leads people to fall in love with a house before evaluating whether the neighborhood truly fits their long-term lifestyle and financial goals.

We were inspired by the gap between what buyers think they need and what actually determines long-term satisfaction and investment return. Research consistently shows that neighborhood factors such as safety, schools, commute time, and economic growth often influence long-term outcomes more than cosmetic house features.

We wanted to build something different, a platform that prioritizes community fit first, and helps buyers evaluate not only the present price of a home, but its future cost and value trajectory.

What it Does

HomeScore flips the traditional real estate search process.

Instead of starting with houses, users rank what matters most to them:

Safety
Property security
Schools
Neighborhood wealth
Commute
Education level
Diversity
Family-friendliness

These preferences form a weighted vector:

$$ P = (w_1, w_2, \dots, w_n) $$

Each neighborhood is represented by its feature vector:

$$ N_i = (x_1, x_2, \dots, x_n) $$

We compute a compatibility score using:

$$ Score_i = P \cdot N_i $$

Neighborhoods are ranked before any house listings are shown.

After a neighborhood is selected, users see houses that match their physical requirements. Finally, we project:

Future home value
Maintenance costs
Best-case and worst-case scenarios

across multiple time horizons (6 months, 1 year, 3 years).
An LLM generates a transparent explanation of why each recommendation was made.

How We Built It

HomeScore follows a four-stage pipeline:

Preference Intake (Frontend)

A React-based interface captures both lifestyle and house requirements.

Neighborhood Classification

We trained a classification model using multi-source data:

Crime statistics
School ratings
Income levels
Commute times
Education demographics

The model ranks neighborhoods uniquely for each user.

Rule-Based House Filtering

Once a neighborhood is selected, hardcoded filters match listings based on beds, baths, square footage, and other physical features.

Regression + LLM Layer

We trained a regression model on historical housing transactions augmented with macroeconomic features such as:

Interest rates
Employment growth
Housing inventory

The projected value function can be simplified as:

$$ \hat{V}_{t+k} = f(V_t, r, g, i, m) $$

where

( V_t ) = current value
( r ) = interest rate trends
( g ) = regional growth
( i ) = inventory levels
( m ) = maintenance lifecycle

If the trained model is unavailable, an LLM fallback generates projections using structured economic inputs.

Challenges We Faced

Collecting clean neighborhood data from multiple sources
Normalizing different datasets into a unified feature schema
Avoiding overfitting with limited historical data
Designing explainable AI outputs instead of black-box scores
Managing time constraints during the hackathon
Balancing frontend polish with backend model development

One major challenge was ensuring that the LLM was used only for narrative explanation, while predictive intelligence remained within trained models.

Accomplishments We’re Proud Of

Successfully inverted the traditional house-first search flow
Built personalized neighborhood classification
Integrated macroeconomic signals into value projections
Designed an explainable scoring pipeline
Delivered a demo-ready React frontend

Most importantly, we created a system where every recommendation is personalized, not static or generic.

What We Learned

Feature engineering often matters more than model complexity
Explainability builds trust in AI-driven financial decisions
Economic forecasting requires uncertainty bounds, not single-point predictions
LLMs are powerful for summarization but should not replace structured predictive models
System design and UX matter just as much as algorithm selection