Current LLM-based workflows either aid developers with granular coding tasks or use free-form prompts to generate full applications (“vibe coding”)—but neither abstracts low-level tasks enough or scales effectively for complex projects. This research introduces a modular design methodology: we decompose software into self-contained modules that communicate only via defined synchronizations. This way, we avoid monolithic generation, simplify maintenance, and exploit common coding patterns learned during LLM pretraining. In this project, we’re going to build tools that embed LLMs in code generation, design a domain-specific synchronization language, and perform case studies and prototype development for both instructional and real-world applications.

With the advent of LLMs, and haven taken several HCI classes at MIT, I found myself constantly reflecting on the future of software. In this SuperUROP, I’m excited to bring my experience both as an HCI enthusiast and as an user of all kinds of AI coding assistance tools into Prof. Jackson’s concept design research.