I’ve done a lot of work in this area for customer interviews. I’ve partnered with a company where I do the AI development, they are building the UI.
A few things in your write up jump out at me based on what we’ve learned:
1. Showing the human what changed from round to round is one of the hardest parts of this problem. We have taken a lot of inspiration from git diff and generate change sets. I wrote a blog post about this on http://producttalk.org.
2. Separating synthesis into two distinct steps help. Step 1: what’s interesting in this interview 2. given these coded interviews, what is interesting across them.
3. We are still exploring how much context to give the agent, but have gotten surprisingly far without having to do much here. I suspect one difference for us is we are working with much longer texts.
Happy to share more of my experience if interested.