Stellantis VPE Innovation Challenge 2025
- Winner: Technical Merit (96% accuracy)
- Judges validated solution on production dataset
- Outperformed competing teams on extraction accuracy
- Lost overall due to presentation execution (learning opportunity)
Data is trapped across formats. Construction companies, lenders, and health platforms receive critical information across PDF specs, Word contracts, Excel BOMs, XML schemas. Extracting structured parameters manually is slow, error-prone, and doesn't scale.
Multi-format document extraction engine achieving 96% accuracy on production datasets.
- 96% accuracy — Judge-validated on real data
- Handles 4 formats — PDF, DOCX, XLSX, XML
- Production-ready — Hybrid NLP + heuristics for reliability
- 1st Place (Technical) — Stellantis VPE Innovation Competition
Pure NLP fails on malformed documents (real-world problem). Regex fallback catches edge cases judges specifically praised for robustness.
These 4 formats cover 95% of real-world business documents.
For high-stakes decisions, users need reliability signals. This was part of what judges valued.
- NLP: SpaCy entity recognition + regex fallback
- Parsing: PyMuPDF, python-docx, openpyxl
- Frontend: Streamlit with filtering, comparison, CSV export
- Validation: Tested on 10K+ documents; 96% accuracy across types
- Clone or download this repository.
- Ensure you have Python 3.7 or higher installed.
- Install the required dependencies using pip:
pip install -r requirements.txt
Run the Streamlit app with the following command:
streamlit run Main.py
This will launch the web interface in your default browser.
- Technical execution matters most — Our solution was strongest; judges confirmed it
- Presentation execution matters too — We learned this the hard way
- Hybrid systems beat pure ML — Judges praised our approach for production reliability
- Iteration > Perfection — Since this project, shipped 5 systems in 12 months
This project is provided as-is without explicit license.