SWE workbench trainer at Turing
At Turing, I worked on the SWE-Workbench AI trainer & DevOps team, building infrastructure and tools for training open-source LLMs on real-world software engineering tasks. Project Description SWE-Workbench is a benchmark/dataset platform that converts GitHub issues and pull requests from popular open-source repositories into executable tasks for training/evaluating AI coding agents. My role on the DevOps team involved: Repository synthesis: Setting up Dockerized environments for 1,000+ diverse codebases (different languages, build systems, dependencies) Task verification: Building pipelines to automatically validate that LLM-generated patches actually pass tests and resolve issues Scalable training infra: Managing the SWE-bench Reviewer (linked) and WMS (Workflow Management System) platforms for human-in-the-loop validation and dataset generation Key links: SWE-bench Reviewer - Task review interface SWE-bench WMS - Workflow management system https://swebench-reviewer.turing.com/1KRB0VB6XxG1lGhhcazlCZ2eAdE7VSktW https://swebench-wms.turing.com/