Stage Deployer

DEMO INSTRUCTIONS:
Test login: judge / FFm6HYpYy/O1LGWb at https://stage.firebots.cloud. The audience side needs no login create an interaction, hit "Deploy to the room", then scan the QR (or click the join URL) from any phone. A fresh AI generation takes ~2 minutes

Framing the Problem

Tech talks are often one-way.
The speaker talks. The audience mostly listens.

People get bored and check their phones.

Normal polling tools are too limited. They need setup before the talk and cannot react to the moment.

Speakers need a way to create live audience activities during the talk.

Idea Explanation

Stage Deployer lets speakers create small audience apps from a simple text prompt.

The speaker types an idea, waits about 2 minutes, then shows a QR code.

The audience scans it and joins right away.
No app. No login. No setup.

Each app matches the talk style.
A Docker talk can look like containers.
A security talk can look like a terminal.

The results appear live on the big screen as charts, rankings, or word clouds.

Implementation

The system has three main parts.

The speaker console is built with React, Vite, and KendoReact.

A Node.js relay sends the prompt to an n8n workflow.
That workflow uses Codex CLI over SSH to generate one HTML file.

The app is saved on a static server and gets its own URL.

The app and console use WebSockets to share live results.

There is no database.
Live data stays in memory.
Old sessions stay in localStorage.
Generated apps are simple files.

The relay has safety limits for many users at once.

Challenges

The biggest challenge was trusting apps that were just created by AI.

I solved this with a strict results format, so the console knows how to read the app data.

Another problem was that my AI credits ran out during the hackathon.
I changed the system to use Codex CLI over SSH.

Performance was also hard.
One visual effect froze my browser, so I banned slow effects in generated apps.

I also lost time because browser caching hid new deploys.
I fixed this with better cache headers and a visible build counter.

Accomplishments

The full flow works today:

Idea to "generated app" to "QR code" to "phones join" to "live results on stage".

I also built:

• a speaker composer

• a real phone preview

• build progress

• interaction history

• AI-suggested interactions

• live stage controls

• 42 unit tests

The biggest lesson was that prompt engineering is also system design.

Next Steps

Next, I want to improve scaling for large rooms.

I also want to add ready-made activities, emoji reactions, and better event support.

Later, I want analytics for organizers and more result types, like heatmaps and live Q&A.

Long term, popular generated apps could become templates that other speakers can reuse.