From Wow to Workflow

How I Generate Music and Deploy Apps with AI Agents

Alexander Gomez · 2026

Let's Start With a Quick Poll

Scan the QR code or go to:

apps.firebots.cloud/20260312150516-a1bess

[ QR CODE ]
Generate at: qr.io or similar

Vote and keep this tab open — we'll come back to it later.

This Is Moving Fast

~100%
of code at Anthropic is now AI-generated — some engineers ship 22 PRs in a single day

4,500
developer-years saved by Amazon's AI tool — $260M/year

$2B
annual revenue for Cursor (AI code editor) — in just 15 months

Goldman Sachs — deploying autonomous AI coders alongside 12,000 developers
Shopify CEO — "Prove why AI can't do it before requesting a new hire"
Google — 30%+ of new code is AI-generated

This isn't just startups. This is enterprise. So I tried it myself.

Anthropic/Fortune Amazon/Digiday Cursor/TechCrunch Goldman/CNBC

Let me set the stage with some numbers. And these aren't predictions — these are facts from the last 12 months.

[Click] At Anthropic — the company behind Claude — nearly 100% of code is now AI-generated. Some engineers ship 22 pull requests in a single day. All written by AI.

[Click] Amazon built an internal AI tool that saved 4,500 developer-years of work. That's $260 million per year. Java upgrades that used to take 50 days? Now done in hours.

[Click] Cursor — an AI-powered code editor — went from zero to $2 billion in annual revenue in just 15 months. The fastest-scaling software company ever.

[Click] And it's not just startups. Goldman Sachs — founded 1869 — is deploying autonomous AI coders alongside 12,000 human developers. Shopify's CEO says: prove why AI can't do it before you hire someone.

[Click] So I thought — let me try this for real. Let me build something serious with AI. And what happened next genuinely surprised me.

EMX

"Turn your moments into cinematic soundtracks"

Upload a photo → AI analyzes the scene → generates lyrics, music, album art, video
Powered by: Gemini (vision + composition), Kie AI (music), Kling (video)
Python/FastAPI backend + Node.js BFF + single-file HTML frontend — all in Docker
Real-time SSE events drive the progressive UI

Photo → Vision → Compose → Image + Music (parallel) → Video

Built in one weekend.

The first project is called EMX. "Turn your moments into cinematic soundtracks."

[Click] You upload a photo — any photo. AI analyzes the scene, the mood, the colors. Then it generates original lyrics, a full music track, an album cover, and even a music video.

[Click] Under the hood: Gemini for vision and composition, Kie AI for music generation, Kling for video.

[Click] The stack: Python FastAPI backend, Node.js BFF, and the frontend? A single HTML file. When the AI suggested a single-file SPA I thought it was being lazy. Turns out it works great.

[Click] The pipeline runs partly in parallel — real-time SSE events keep the UI updated as each step completes.

[Click] And here's the thing: I built this in one weekend. With an AI agent.

Let me show you.

LIVE DEMO

EMX

emx.firebots.cloud

Emotion Pad — 2D valence/arousal input
Upload a photo, watch face detection move the dot
Walk through the pipeline as it happens
Album cover, music, video — all generated live

How EMX Was Built

One weekend. One AI agent. Five services.

Day 1: FastAPI backend + pipeline orchestration
Vision analysis, lyrics composition, parallel music + image generation
Day 2: Frontend + BFF + real-time SSE
Emotion Pad, face detection, progress animations, waveform player
The honest part: I iterated for a second weekend on polish
Gallery, bug fixes, parallel pipeline optimization, dev mode for testing

The AI wrote the code. I made the architecture decisions, reviewed every line, and debugged the parts it got wrong.

So how was this actually built? Because the abstract promises I'll show you — and I will.

[Click] Day 1: I described the pipeline to the AI agent. "I need a FastAPI backend that takes a photo, analyzes it with Gemini, composes lyrics, generates music and an album image in parallel." It wrote the entire backend. I reviewed it, made architecture calls, fixed a few things.

[Click] Day 2: Frontend and the BFF layer. The Emotion Pad, face detection using face-api.js, progress animations — all generated by the AI. I guided the design, caught bugs, made it feel right.

[Click] Honestly? I spent a second weekend polishing. Gallery view, bug fixes, making the parallel pipeline actually reliable. The AI is fast but it's not perfect.

[Click] And that's the key insight: the AI wrote the code. I made the decisions. I reviewed every line. I debugged what it got wrong. That's the workflow.

Vibe Deployer

"Describe an app → AI builds it → deployed in 90 seconds"

Chat interface → your prompt hits an n8n workflow → Claude API writes the app
WebSocket relay pushes real-time progress back to the browser
Auto-deployed to apps.firebots.cloud/{slug}

Four containers. One prompt. Deployed app.

Also built in one weekend.

LIVE DEMO

Vibe Deployer

vibe.firebots.cloud

Type a prompt — something fun
Watch it generate in real-time
Open the deployed app

REVEAL

Remember that poll you voted on at the start?
That was built with the Vibe Deployer. One prompt. Under 90 seconds.

Let's build something live. What should we make? Something fun — games work great, visual stuff.

[Take a suggestion from the audience or have a backup ready.]

I'll type it in... and now we watch.

[Narrate the WebSocket updates as Claude writes the code.]

You can see it generating... the HTML, the CSS, the JavaScript...

[When deployed, open the app.]

There it is. A working app, deployed, with a URL you can open right now on your phone.

[Pause. Then the reveal.]

Now — remember that poll you voted on at the very start of this talk?

[Click to show the reveal box.]

That was built with the Vibe Deployer. One prompt. Under 90 seconds. You've been using an AI-generated app this whole time.

[Let it sink in. This is your biggest wow moment.]

How I Actually Work

The tool: Claude Code — not a chatbot, not an IDE plugin. An autonomous agent in your terminal.

Full codebase awareness — reads, searches, edits files, runs commands
Plan mode: Claude proposes a step-by-step plan. I review. THEN it executes.
Runs on my Linux server via SSH — commits → GitHub Actions → auto-deploys
CLAUDE.md as persistent memory — it remembers my server, my conventions, my rules
I work from anywhere — phone, laptop, any terminal

Armin Ronacher — Creator of Flask & Jinja2

"Review every line, shape the architecture, and carry responsibility."

Ronacher/lucumr.pocoo.org

So what's behind all this? The tool I use is called Claude Code.

It's not ChatGPT. It's not Copilot in your editor. It's an autonomous agent that lives in your terminal.

[Click] It reads your entire codebase — not just one file. It searches, edits, runs commands.

[Click] The key: plan mode. I describe what I want. Claude reads the code, proposes a plan. I review it — "yes," "change this," "no, do it differently." Only then does it write code. Measure twice, cut once.

[Click] It runs on my server. When it commits, GitHub Actions auto-deploys. The cycle is: I talk to Claude, it writes, it commits, it deploys. All from the terminal.

[Click] CLAUDE.md — a memory file. It knows my Docker setup, my routing rules, what volumes not to delete. It persists across sessions.

[Click] I work from anywhere. Phone on the couch. Laptop at a cafe. Any terminal with SSH.

[Click] Armin Ronacher — the creator of Flask — said it best: "Review every line, shape the architecture, and carry responsibility." That's exactly how I work.

The Three Levels

Level	How	Who Drives	Example
1. Chat	Copy-paste prompts, manual integration	You type everything	ChatGPT, Claude.ai
2. IDE	AI in your editor, inline suggestions	You + AI side by side	Cursor, Copilot
3. Agentic	AI on your machine, autonomous execution	AI executes, you review	Claude Code, Codex

Level 1 is talking to a tutor. Level 2 is pair programming.
Level 3 is having a developer on your team that you manage.

Remember the poll? Check your results — where does this room land?

What I Learned Along the Way

Plan mode is non-negotiable
"Measure twice, cut once" applies to AI too
You must understand the code
You're the architect, not the AI
The AI makes mistakes
But so do humans — and AI is faster at fixing them
It's not about the prompt — it's about the review
Knowing what to accept, what to reject, and what to ask again

Kent Beck — Creator of TDD & Extreme Programming

"90% of my skills dropped to $0. The leverage for the remaining 10% went up 1000x."

Kent Beck/Substack

Four things I learned the hard way.

[Click] One: plan mode. Non-negotiable. Every time I skip it, I regret it. Let the AI think before it acts.

[Click] Two: you must understand the code. The AI is the builder. You're the architect. If you can't read what it wrote, you can't catch what it got wrong.

[Click] Three: it makes mistakes. It generates bugs. It misunderstands requirements. But honestly — so do humans. And the AI doesn't send passive-aggressive Slack messages at 3 AM about it.

[Click] Four: it's not about the prompt — it's about the review. The skill is knowing what to accept, what to reject, and when to say "try a different approach."

[Click] Kent Beck — the inventor of Test-Driven Development — put it perfectly: "90% of my skills dropped to zero. The remaining 10% went up 1000x." The game has changed. But you still need to know the game.

My Challenge to You

The question isn't whether AI can help you.
The question is: at which level are you going to engage?

Try an AI agent for one week. Use plan mode. Review every line.
I guarantee you won't go back.

This entire talk — the slides, the speaker notes, the research — was built with an AI agent on my server. I mostly supervised from the couch.

Questions?

emx.firebots.cloud · vibe.firebots.cloud · claude.ai/code