Hey everyone! 👋 I’ve been working on Adeptly (an interactive, adaptive DAP). You can see a quick demo here: • Originally, I was over-engineering everything by calling OpenAI Realtime API and Gemini directly. It was a mess. I decided to rebuild from scratch using LiveKit, and in just 1-2 days, I had a much cleaner, production-ready product. However, I hit a new wall: Real-time Prompt Engineering. Manual testing was frustrating and slow. I could get to ~80% accuracy, but hitting 90%+ felt impossible. I eventually switched to automated integration testing (using text mode) and discovered GEPA (Automatic Prompt Optimization): • https://github.com/gepa-ai/gepa This changed everything. By shifting my focus to building high-quality datasets and letting APO handle the prompting, I reached 98% accuracy. That experience inspired me to build Zelta. It’s a tool designed to help build and iterate on those crucial datasets (especially synthetic ones) to make optimization easier. I use it internally for my own product, but I wanted to share it here in case it helps you streamline your own prompt engineering workflow. Feel free to check it out: • https://zelta-chi.vercel.app/

That experience inspired me to build Zelta. It’s a tool designed to help build and iterate on those crucial datasets (especially synthetic ones) to make optimization easier.

Cool, I have seen a couple of people say they need this kind of automated prompt optimization recently

Yeah.. manual prompt engineering drive me crazy.. so I think this dataset-driven automatic prompt optimization could be one of the alternatives.. btw.. if you mind to share.. do you have the same of frustration as I am? @lively-winter-88668

really cool! are there any privacy concerns on scren capture capturing sensitive info? and also, that green bubble waveform is great! is that a livekit component?

@powerful-magician-63563 don't think that far yet about privacy concerns. Still focus on validating the idea by shipping it to real users. The green bubble is basically copying from Google Live API official demo 😂, just changed the color and some other things a little bit