parassss17/homluv-preference-engine

HomLuv — Visual Home Preference Engine (Prototype)

A working prototype of HomLuv's core idea, built during an industrial internship at EX Squared Solutions (a Builders Digital Experience / BDX company).

HomLuv is a visual search & discovery platform for new-home buyers: instead of typing filters, buyers like / dislike home photos, and the platform learns their style and matches them to the right homebuilder by price and location. This repo reproduces that loop end-to-end on a small public sample dataset, so the technique is demonstrable without any private data.

Scope note (read this): This is a prototype on the public Houses Dataset, not HomLuv's production system. The dataset has no builder field, so builder portfolios here are simulated by clustering homes on visual style + price. All numbers in the notebooks come from this code on this sample — nothing is taken from the company.

The idea in one line

Average the image vectors of the homes a buyer liked into a single "taste" vector, then rank every home by how close it is to that taste — and suggest the builder whose homes match best, within budget.

Pipeline

data/Houses-dataset  (535 homes: 4 photos each + price/zip)
        |
        v
prepare_data.py   (run ONCE, locally)
   - photos --> 1280-dim style vectors   [pretrained MobileNetV2, NOT trained by us]
   - KMeans on style+price --> simulated "builders"
        |
        v
artifacts/   embeddings.npy  +  listings.csv  +  images_sample/   <-- small, committed
        |
        +--> notebooks/01_explore   (EDA)
        +--> notebooks/02_model     (recommender demo + classifier accuracy)
        |
        v
app.py   Streamlit demo: heart photos --> taste vector --> matches + builder

The heavy image model runs only in prepare_data.py on your machine. The deployed app loads just the small cached vectors (NumPy / scikit-learn), so it runs on free Streamlit Community Cloud or Hugging Face Spaces.

What each library does

Run it locally

# 1. get the data (see data/README.md), then:
pip install -r requirements-dev.txt
python prepare_data.py          # one-time: builds artifacts/

# 2. explore / verify
jupyter notebook notebooks/

# 3. run the demo
streamlit run app.py

Deployed app needs only requirements.txt (no PyTorch).

Author

Paras Beniwal — built during the EX Squared Solutions internship (Jun–Aug 2025), on the HomLuv project.