DEV Community

Crystal Chang
Crystal Chang

Posted on Edited on

What Building Zukki Taught Us About Honest AI Demos

We’re building Zukki, a night-shift assistant for small local-service businesses, starting with plumbers. A customer can describe a problem, while the business owner reviews the request and decides what happens next.

As we prepared our hackathon demo, we ran into a question: how do we show what we built without making every part of the demo look live?

Here’s what we learned.

A real channel test and a simulator tell different stories

We tested a real WhatsApp conversation on November 1. We also built a separate local demo that uses a WhatsApp simulator and recorded model responses. The simulator makes a scripted flow easier to reproduce, but it doesn’t prove that every step happened through WhatsApp.

Our solution was simple: label each part where it appears, rather than asking viewers to work it out afterward.

Recorded model outputs can make a demo reproducible

Our local demo replays recorded model responses, so judges can try the scripted flow without providing their own API keys. We also have separate footage of a November 3 run using live models. Both are useful, as long as we’re clear about which is which.

One limitation: the no-key replay supports the recorded scenario, not every possible customer request.

Show the human decision, not just the AI reply

The customer conversation is only half of the story. We also wanted to show what the owner sees: reviewing a request, preparing a quote, and deciding what to do next.

For our Story Pack demo, the records are fictional, and the video makes clear that no post was sent. That boundary matters to us.

A small checklist we’ll keep using

  • Was this step live, simulated, or replayed?
  • What actually happened in this run?
  • What did a person review or approve?
  • Which data is fictional?
  • Can someone else reproduce the shown flow?

See Zukki in action

AI assisted with drafting this post.

Top comments (0)