How to Design and Deploy the Best Vision Pipeline

​When you're designing a system that can see and understand the physical world, you want to find the setup that performs best. That might mean comparing hundreds of possible configurations: different models, different prompts, and different logic. Testing each combination by hand requires saving every output and eyeballing the results, which takes time and might give you unclear answers.

​In this live session, we'll look at two tools that help you run those experiments and find the optimal configuration. We'll start with a project where we needed visual intelligence to watch a video and answer various questions about it. That example will show how you can explore different configurations and find the one with the best accuracy and latency.

​What you'll see:

• ​Compare configurations side by side: Pick a set of test images, add the right answers, run many versions of a Workflow against them, and see accuracy and speed for each one in one view.

• ​Explore vision pipelines automatically: Have Roboflow Agent build the variations, or use Claude Code, Codex, or Cursor to build them, test them, and report back.

• ​Search thousands of model designs: Train once while Neural Architecture Search (NAS) tests over 5,000 model configurations, then pick the best model at each speed for the hardware you deploy on.

​By the end of the session, you'll know how to set up a Workflow Eval and interpret its results, and how to use NAS to choose a model that balances accuracy and speed.