Bridging Code and Context: How Knowledge Shapes AI Software Engineering

Title of the Talk: Bridging Code and Context: How Knowledge Shapes AI Software Engineering
Host Faculty: Dr.Anupam Sanghi
Speaker: Mr.Rohith Pudari
Date: 13 August
Time: 11:00 am
Venue: CSE Seminar Hall

Abstract
Large language models and coding agents are changing software engineering by helping developers write code, modify existing software, and solve programming tasks. However, these systems are still evaluated on whether they produce working code, not on whether they understand the software they change. Software engineering decisions depend on context beyond source code, including documentation, project history, and domain knowledge. My research studies how this context affects AI systems that perform software engineering tasks.

I investigate this problem through scientific software, where correctness often depends on domain assumptions that are rarely captured in code. In this talk, I present three findings. First, an empirical study of refactoring detection, showing that existing tools miss software changes that require domain understanding. Second, interviews with scientists reveal how developers use domain expertise, project history, and collaboration when making decisions that current tools cannot support. Third, I develop a taxonomy of knowledge required for software engineering tasks and evaluate how providing this knowledge affects autonomous coding agents. Providing relevant knowledge improves task success and reduces unnecessary exploration, even when agents fail. However, agents still spend significant computation searching and using context inefficiently.

These studies show that effective AI software engineering requires the knowledge developers use to understand software and make decisions. However, providing more knowledge alone is insufficient. My future research focuses on making knowledge use more efficient by measuring how agents spend computation, identifying unnecessary actions, and developing methods that help agents select relevant context without sacrificing task success. My goal is to help build AI systems that use context efficiently to understand software and make reliable changes.

Bio

Rohith is a final-year Ph.D. candidate at the University of Toronto, advised by Prof. Shurui Zhou. His research lies at the intersection of software engineering and artificial intelligence, with a focus on building knowledge-aware AI systems that reason about software beyond source code. He holds an M.Sc. in Computer Science from the University of Victoria, advised by Prof. Neil Ernst, and has interned as a Research Software Engineer at Meta, where he built a trajectory debugger for LLM agents that cut iteration time on agent failures by more than 50%. Beyond academia, Rohith has made notable open-source contributions to widely used machine learning ecosystems, including Google’s Keras and OpenMined’s PySyft. More information is available at https://rohithpudari.github.io/