Natural language layout generation with consistent styling
Webflow is building the world’s most powerful website experience platform to bring development superpowers to everyone.
To help website builders automate repetitive tasks and streamline site build workflows, Webflow began development of their AI Assistant – a natural language assistant which understands the context of the site, its styles and content, and will automatically tailor new sections to match.
As Webflow developed the AI Assistant, they ran into issues with reliability and regressions as they added new capabilities, like the ability to ask follow ups. As such, they started looking for a system to implement evaluations.
Multimodal, team-wide grading
Webflow chose Gentrace as their evaluation system, implementing a system that used a combination of model, human, and code-based evaluation techniques.
In particular, Webflow valued how Gentrace:
- Supported multi-modal (image) outputs, including side-by-side comparison and LLM-as-a-judge grading
- Allowed over 25 stakeholders, including product managers, engineers, and leadership, to gain visibility and contribute to evaluation.
- Increased testing breadth and depth from ad-hoc analyses to thousands of evaluations per day.
Bryant Chou, co-founder and chief architect at Webflow, shared:
Every LLM product needs evals. Gentrace makes evals a team sport at Webflow. With support for multimodal outputs and running experiments, Gentrace is an essential part of our AI engineering stack. Gentrace helps us bring product and engineering teams together for last-mile tuning so we can ship AI features that delight our users.
A successful launch of AI Assistant
With a system for testing and the input of many stakeholders, Webflow successfully launched their AI Assistant this past October at their Webflow Conference. We’re thrilled to team up with Webflow to see how AI Assistant and their other AI products evolve to supercharge website creators moving forward.
