Project 20 - Multi-Modal ERD (Tables + Images)

Graduate Mentor: Enamul Haque

Graduate mentor's supervisor: Grant Weddell

Research Problem: 

Data lakes contain massive amounts of structured tables and unstructured images. Currently, database systems cannot connect them automatically; a table of store products and a folder of product images remain completely isolated. This research bridges that gap using vision-language AI models to map text, data columns, and images into a shared space. Solving this unlocks "dark data," allowing organizations to fully search, analyze, and use all their visual and text data together for the first time.

Importance:

Instead of rigid, exact matches (like matching ID numbers), you will build a Probabilistic Entity-Relationship Diagram (pERD). This AI-driven system connects tables and images based on concepts and visual similarity (A connects to B with 85% confidence). It’s an exciting mix of databases and modern multimodal AI!

Working in a team of 3–4, you will split up tasks like:

  • Short-term: Use pre-trained AI models to generate embeddings for a dataset of tables and images.

  • Medium-term: Build a graph structure to map and visualize these probabilistic AI links.

  • Longer-term: Create a simple interface to let users review and refine the AI's connections.