Kitana Agentic Wrapper

TLDR

• Built an agentic wrapper on top of the Kitana AutoML system to intelligently select and clean joinable tables under a resource budget.
• Designed and tested multiple agentic architectures containing embedding-based pruning and LLM-powered reasoning for table ranking and filtering.
• Achieved 12.5% higher R² than embedding-only baselines on synthetic table benchmarks — supporting both no-hop and multi-hop joins.
• Implemented full pipeline in Python with OpenAI and Gemini API calls for LLM reasoning and embedding generation.

Project Summary

As data lakes continue to grow, ML engineers face a painful bottleneck: identifying the right tables to clean and join to improve model performance. This project extends Columbia’s Kitana AutoML system, which searches for useful joins to improve a target column’s prediction accuracy.

We proposed an agentic system to decide which tables to clean — under a budget — using past Kitana queries, accuracy improvements, and semantic cues. Our system beats a naive embedding baseline by up to 12.5% in R² and supports both no-hop and multi-hop table selection.

Architecture: Selector Agent

Selector Agent: Embedding-based pruning + LLM-based enrichment pipeline

Code & Report

What I Learned

Contributions and Acknowledgements

This project was completed as part of COMS 6113: Agentic Systems Made Real, a graduate-level research seminar taught by Professor Eugene Wu at Columbia University. I worked alongside Mateo Juliani (msj2164@columbia.edu) and Kaushal Damani (akd2990@columbia.edu).