Machine Learning & AI
SpaceWorth
SpaceWorth combines a custom-trained 90.64% R² Machine Learning price valuation ensemble (LightGBM + CatBoost + 3x PyTorch Entity Embedding Neural Networks) with multimodal CAD floor-plan OCR parsing (Gemini Flash) to deliver instant, verifiable real estate estimates.
SpaceWorth — AI Real Estate Valuation & CAD Intelligence
ITI Machine Learning & AI Engineering Final Project
Status: Live
The Problem
Traditional real estate pricing relies on manual inspections or simple linear heuristics that fail to capture location dynamics and architectural layout nuance. Converting engineering drawings, blueprints, or CAD floor-plan PDFs into valuation inputs requires tedious manual measurement.
The Solution
SpaceWorth combines a custom-trained 90.64% R² Machine Learning price valuation ensemble (LightGBM + CatBoost + 3x PyTorch Entity Embedding Neural Networks) with multimodal CAD floor-plan OCR parsing (Gemini Flash) to deliver instant, verifiable real estate estimates.
What Does SpaceWorth Provide?
- 90.64% R² Custom ML Model Engine: Trained from scratch on 57,058 Indian real estate listings. Combines target encoding, LightGBM, CatBoost, and 3x PyTorch Entity Embedding NNs hosted on Hugging Face.
- CAD Floor-Plan Intelligence: Upload blueprints, CAD drawings, or architectural PDFs (PNG, JPG, PDF up to 12MB). Gemini extracts room counts, dimensions, and sqft directly into our valuation engine.
- Interactive 2D/3D Konva Canvas: Dynamic floor-plan layout generator allowing users to add, drag, resize, measure, and preview room configurations in 2D and 3D.
- Developer API Key Manager: Full API key lifecycle management stored in Supabase PostgreSQL with secret hashing and rate limiting.
Model Performance & Benchmarks
- • Blended Ensemble (Selected): 90.64% R² Accuracy
- • CatBoost Model: 88.40% R²
- • PyTorch Entity Embedding NN: 86.90% R²
- • LightGBM Model: 85.12% R²
- • Ridge Regression Baseline: 79.59% R²