Advancing AI requires far more than simply building smarter language models. On their own, models face clear limits: they do not naturally optimize solution strategies under real-world constraints like latency or cost, they struggle to reuse knowledge across tasks, and they often execute plans inefficiently, especially when coordinating across multiple models and external tools. My research focuses on designing system and software infrastructure that surrounds and augments language models, enabling AI to optimize for practical, multi-objective goals such as accuracy, efficiency, and budget.

My research also involves building datasets and benchmarks that evaluate AI performance, including BEAVER for enterprise Text-to-SQL and MDCR for multi-document conditional reasoning.