Predicts home value from property and location data, then generates negotiation strategy grounded in real comparable sales and market conditions.
Buyers and sellers rarely have access to the data-driven pricing and negotiation insight that real estate agents use internally. ValueIQ estimates home value using machine learning, grounds that estimate in real neighborhood and market context via retrieval, and generates negotiation strategy from an LLM reasoning over that context — not just generic advice.
- Estimates home value from property attributes (square footage, bedrooms, lot size, year built) and location (neighborhood, kitchen quality)
- Surfaces 5 comparable sales retrieved semantically from a vector store of 2,927 embedded property records
- Generates negotiation strategy grounded in comps and ML prediction rather than generic tips
- Color-coded valuation summary — green when asking price is below predicted value, red when above
The system has four stages:
- Data layer — Ames Housing dataset, cleaned and feature-engineered across 6 notebooks
- Price prediction model — Ridge regression (α=10) trained on 211 features after one-hot encoding
- RAG context retrieval — FAISS vector store with 2,927 embedded property documents using sentence-transformers (all-MiniLM-L6-v2), returning top-5 comparable sales
- LLM negotiation engine — Google Gemini generates strategy grounded in the retrieved comps and model output
- Modeling: Python, scikit-learn, XGBoost, pandas, numpy
- RAG: FAISS, sentence-transformers (all-MiniLM-L6-v2)
- LLM: Google Gemini API (gemini-2.5-flash)
- Backend: FastAPI, uvicorn
- Frontend: React 18, Vite, react-markdown
- Data: Ames Iowa Housing dataset via Kaggle
This project uses publicly available, aggregated datasets — it does not scrape or use live Zillow listings or Zestimate data.
- Housing data: Ames Iowa Housing Data via Kaggle
Predictions are based on Ames, Iowa housing data and are not a substitute for professional appraisal.
git clone https://github.com/Joaco273/ValueIQ.git
cd ValueIQ
# Backend
pip install -r requirements.txt
cd src
uvicorn api:app
# Frontend (separate terminal)
cd frontend
npm install
npm run devCreate a .env file in the project root:
GEMINI_API_KEY=your-key-here
Run notebooks 01-06 in order to regenerate the model and vector store before starting the backend.
| Model | RMSE | MAE | R² | $ RMSE | $ MAE |
|---|---|---|---|---|---|
| Linear Regression (baseline) | 0.0985 | 0.0737 | 0.9460 | $18,414 | $12,999 |
| XGBoost (tuned) | 0.1092 | 0.0800 | 0.9337 | $24,429 | $15,011 |
| Ridge (α=10) — final | 0.0950 | 0.0716 | 0.9497 | $18,157 | $12,767 |
Ridge outperformed XGBoost on this dataset — strong linear relationships and clean feature engineering meant regularized linear regression captured most of the signal without the added complexity of tree-based methods.
- Phase 1 — Clean dataset, train and evaluate price model
- Phase 2 — Build RAG layer for neighborhood/school context
- Phase 3 — Add LLM negotiation reasoning grounded in comps
- Phase 4 — Polish frontend, deploy live
MIT — see LICENSE for details.