Bijay Yerra

Repair Relay, a tool that helps people directing AI coding assistants verify whether a bug fix actually meets their requirements. The idea comes from my own AI coding experiments: an application could build successfully and pass its component tests while still failing basic user interactions. Repair Relay will use Meta models to attempt repairs, guided by independent verification. A first model proposes a fix, then a separate runner executes acceptance tests—including user workflows—that the model cannot modify. If verification fails, a second Meta model receives the attempted patch and failure evidence for another attempt. For the ten-day hackathon, I will focus on three small Python projects and a working interface showing changes, test results, elapsed time, and model usage. I will compare the relay with a single-model baseline and report both successes and failures. The goal is to make the AI-generated repairs easier to evaluate through observable evidence.