Six months ago I couldn't finish a Udemy course. Last week I deployed a multi-agent AI system to production.
This isn't a flex. It's a story about what's actually possible when you stop waiting until you feel "ready."
🚀 How It Started: A Sold-Out Mac Mini
My first real exposure to AI wasn't a course or a bootcamp.
It was watching the AI agent meme coin craze on Solana. I watched a creator launch the first LLM-based meme coin and thought — this is something different.
I'd dabbled with coding once, during the pandemic:
- A Udemy course I never finished
- One project I can't even remember
- That was the extent of my technical background
But the YouTube algorithm kept feeding me videos about building local AI agents. I watched one. Then another. Something clicked.
I went to Best Buy to buy a Mac Mini to try it myself.
They were sold out. Every single one.
I drove home and ordered one from Apple online instead.
When a product sells out in every Best Buy in the city, it usually means something real is happening.
A month later my Mac Mini arrived. That's when everything changed.
🤖 Cerberus: Building an Autonomous Trading Bot From Zero
My first project was Cerberus — a fully autonomous AI trading system built on a Mac Mini M4 using n8n, Claude, and the Solana blockchain.
I didn't start from scratch. I found an n8n community template called "AI-powered stock analysis assistant with Telegram, Claude & GPT-4O Vision" and tried to get it working.
It didn't work.
- The Merge node was broken
- The webhook kept failing
- Basic commands like "analyze AAPL" returned nothing
I spent hours on a single node trying different modes — Choose Branch, Combine, Append — before finally landing on the right one.
But what started as "get this template working" eventually became something I didn't expect to build:
What Cerberus Could Do:
✅ Live AI stock analysis with charts sent directly to Telegram
✅ Real Solana token swaps via Jupiter
✅ Daily 9am morning briefing covering 5 watchlist stocks
✅ Market scanner finding top gainers in real time
✅ Persistent memory via Supabase so the bot remembered every conversation
✅ Live trading dashboard with a Midnight aesthetic
A Telegram bot called Cerberus that could analyze stocks, execute real trades, and brief me every morning — all from a single chat interface.
The Obstacles Were Relentless
The worst was webhook instability. My Cloudflare tunnel URL kept changing every time n8n restarted, dropping the Telegram connection.
This wasn't a one-time fix — it came back every single session.
We tried:
- Environment variables
-
.zshrcconfigurations - 30-second delays, 60-second delays
- Cron jobs running every 5 minutes
Nothing stuck permanently.
Then there were n8n routing bugs — If node conditions behaving inconsistently, requiring hours of debugging to isolate.
Alpha Vantage's 25 requests-per-day limit constantly blocked features I'd just built.
Making changes without backups caused repeated regressions that set me back hours.
At one point a startup script I added to simplify things caused more problems than the three separate terminals it was supposed to replace.
I said out loud:
"I'm regretting doing this at all when it was working just fine before."
But I never once thought about quitting. My newfound excitement for this technology wouldn't allow it.
💡 What Cerberus Taught Me:
- Back up before every change — no exceptions
- Test one thing at a time or you won't know what broke
- Free API tiers have real limits that affect production systems
- Agentic AI development requires a developer mindset, not a freestyle approach
- No version control is a trap — every change without Git is a risk
🧠 The Shift: From Making It Work to Understanding It
After Cerberus I felt something I hadn't expected — I wanted to actually understand what I'd built.
n8n is a visual tool. Powerful, but visual. You connect nodes and workflows without necessarily knowing what's happening underneath.
I wanted to write real code. I wanted to know why things worked, not just that they did.
So I started a new project: a Personal AI Research Assistant.
- Pure Python
- No visual tools
- Three Claude agents running in parallel:
- Search Agent 🔍
- Analysis Agent 🧪
- Report Agent 📄
Each with a specific job. Each building on the previous one's output.
This is where I learned what I actually didn't know.
🐛 What Real Debugging Looks Like
The first error that stopped me cold: my API key wasn't loading.
I stared at the same four lines of code for an hour.
The fix? My .env file was in the wrong folder — one directory up from where the server was running.
Then CORS. My React frontend couldn't talk to my FastAPI backend because I hadn't added the right port to the allowed origins list.
A one-line fix that took 45 minutes to find.
Then the server kept crashing when I added new files because of a Python version incompatibility — str | None type hint syntax doesn't work in Python 3.9.
None of these were glamorous problems. They were the kind of errors that make you feel stupid.
But every single one taught me something about how these systems actually work — not just how to copy them from a tutorial.
⚡ 48 Hours, Three Projects
By the end of two days I had shipped:
1️⃣ AI Research Assistant
Three Claude agents that research any topic in parallel and return a structured report.
🚀 Deployed live with authentication and rate limiting at ai-research-assistant-snowy-pi.vercel.app
- FastAPI backend on Railway
- React frontend on Vercel
2️⃣ AI Memory Agent
A personal AI assistant that remembers you across conversations using RAG and vector embeddings stored in Supabase pgvector.
Every conversation retrieves relevant memories and injects them as context before Claude responds.
3️⃣ AI Tool Agent
An autonomous agent with real tool calling:
- 🔍 Searches the web via Brave API
- 🌐 Reads any URL
- 🐍 Runs Python code
- 🐙 Interacts with GitHub
- 📚 Searches Wikipedia
- 🧠 Manages its own long-term memory
It plans complex tasks before executing them and evaluates its own responses after every interaction.
Three GitHub repos. One live deployment. All built with no computer science degree, no bootcamp, no formal background.
💬 What I'd Tell Someone Who Feels Like They Can't Do This
The most common thing that stops people isn't lack of ability. It's comparison.
You look at:
- Job descriptions requiring CS degrees and five years of experience
- Engineers on Twitter who seem to have been born knowing this stuff
- A line you haven't crossed and maybe can't
But AI has changed what's possible for self-taught builders.
Not because AI writes the code for you — it doesn't, not really.
Because AI can be your teacher, your debugger, your rubber duck, your senior engineer, and your code reviewer all at once.
I didn't follow a curriculum. I built something I wanted to exist.
- When it broke, I debugged it
- When I didn't understand something, I asked until I did
- When it worked, I shipped it and moved to the next thing
That's the whole method.
Pick something you want to build. Build it. Break it. Fix it. Ship it. Repeat.
The background doesn't matter as much as the reps.
The best way to learn AI engineering is to build AI systems.
So go build something.
🎯 What's Next
What I'm building next: Gmail/Calendar integration for my tool agent and scaling the system to handle rate limits better.
Your turn: What stopped you from starting your first AI project? Drop a comment — I read and reply to all of them.
👉 Follow me for Part 2 where I'll break down the code architecture and show you how to build your own multi-agent system.
- 💻 GitHub: [@ivancazares2k]
- https://github.com/ivancazares2k
Top comments (1)
👏