Talk
ML Without the YAML: Replacing Configuration Glue Code with Pure Python
AI has made experimental, iterative workflows everyone's problem. Every software engineer testing prompts, tweaking agentic workflows, or swapping out model providers now lives in the same cycle of "change config, run, eval, repeat." And yet we're still managing that iteration with YAML files, string-based class dispatch, and glue code that our IDEs can't follow. This talk introduces confingy, an open-source library that replaces all of that with plain Python. A single @track decorator lets you track and configure arbitrary Python classes with serialization, validation, and lazy instantiation as a beneficial byproduct — no config schemas required. Through confingy, we were able to pull off a full YAML-killing refactor in 5 months. I’ll cover what worked, how this refactor impacted our software architecture, and what’s still a challenge to this day.
About
Ethan Rosenthal is a Member of Technical Staff at Runway, an applied AI research company focused on multimedia content creation, where he builds engineering systems to accelerate the work of research scientists. His career spans diverse roles across AI, machine learning, and data science - from training language models at Square to developing recommendation systems at seed-stage ecommerce startups. Before working in tech, Ethan was an actual scientist and got his PhD in experimental physics from Columbia University.