From Guesswork to Greatness in LLM Development 

**Abstract** LLM apps hide many ungoverned choices—prompts, retrieval, routing, decoding—making systems brittle and costly. We ask whether these apps are underspecified and show how to replace guesswork with optimization. The presentation and demos uncovers hidden parameter interactions, reads Pareto trade-offs (accuracy–cost–latency), and uses budget-aware exploration to find strong configurations with far fewer trials. We briefly situate these results in a specification-driven workflow for evaluation-driven agent development. **About the Speaker** Dr. Nimrod Busany is the founder of Traigent and a Senior Research Scientist. He previously led an applied AI research team at Accenture Labs and spent nearly seven years at IBM Research–Israel (Haifa & Givatayim). His work bridges knowledge graphs, ontologies, and LLM systems to improve reliability and efficiency of LLM applications; he holds a Ph.D. from Tel Aviv University, is an inventor on multiple patents with publications, and serves on program committees for leading software-engineering venues.