Surveys are suck for two reasons: they are boring to fill out and responses tend to be shallow. We've created the first conversational survey which asks follow-up questions like a human would. It's an AI-led user interview.
The problem which an unstructured interview is that the data is hard to analyze. It takes days to extract insights from thousands of open-ended responses.
Therefore, we built an AI engine that will automatically cluster and summarize data from the survey. The 100k context from Claude was a game-changer in developing this tool.
Technical challenges
It was challenging to cluster the data correctly. We used Claude with multiple steps + KNNs with embeddings to increase the accuracy of the clusters. This way we can give Claude feedback to provide better clustering. This was based a on new paper.
Built With
- claude
- nextjs
- typescript
Log in or sign up for Devpost to join the conversation.