Jarvis: How Mario De Luca Is Building His Own AI Clone

The moment it clicked

The idea for a clone of my own didn't come from a movie, it came from watching something. I saw an AI build a website in a few minutes. Not a draft, not a sketch, but a page you could actually show someone. My first thought was: this isn't a tool. This is an employee.

I offered that first "clone" to businesses back then, and that's where my first clients came from. Before that, at 18, I'd had four attempts that all went nowhere – the story is in my piece on those four failed attempts. The difference this time: I wasn't trying out a new tactic, I was building a system that actually gets work done.

Jarvis is the logical next step of that thought. Not a clone for one task, but one that goes through the day with me.

What Jarvis can do today

Build websites

That's where it started, and it's still one of its strengths. I describe what a page needs to do, Jarvis builds a first version, I review it and give corrections. The bottleneck is no longer the building; it's knowing what makes a good page.

Find leads

Jarvis looks for businesses that fit an offer and collects what you should know before the first contact. Who actually gets contacted is my call. A machine can sort, but it doesn't know who it feels right with.

Write outreach

Jarvis drafts the first message and the follow-up. Drafts, not automatic sends. And one thing matters here: a machine doesn't change the law. If you send marketing emails without consent, you have the same problem with a template as without one. What I know about this from practice is in AI in sales (German) – as a working basis. This is not legal advice; rules differ by country (in Germany: UWG/GDPR).

Work through to-dos

I give Jarvis a task and it works through it as far as it can. Where it gets stuck, it says so. That sounds trivial, but it's the real value: a clone that admits where it's stuck is more useful than one that pretends everything is done.

Voice control

I can talk to Jarvis instead of typing. That's the part that looks most like Iron Man, and at the same time the part that changes everyday life the most: an idea I say out loud while walking ends up as a task instead of getting lost.

How an AI clone works, in principle

A clone isn't a single program, it's a chain: input, understanding, execution, result. The input is voice or text. The understanding is handled by a language model – in my case Claude (more in Claude for business, German). Execution happens through tools the clone is allowed to use: files, a browser, code. And at the end there's a result that a human checks.

The difference between a chat and a clone lies in exactly those tools. A chat answers. A clone acts. What an agent like that can do and where it hits its limits, I've written up in AI agents for business (German). The building blocks you need are in AI tools for business (German).

Why I build it myself instead of using an off-the-shelf assistant

There are ready-made assistants, and many of them are good. I still build mine myself, for three reasons. First, I understand every part. When Jarvis gets something wrong, I know which link in the chain to look at instead of waiting for an update. Second, it adapts to my work and not the other way round: which tools it may use, which tasks it takes on and where it stops are my decisions. Third, building it teaches me exactly what I later show others. A clone you only use stays a black box. A clone you built yourself becomes a craft you can pass on.

The price is time. Building it yourself means things break that someone else would have fixed in a finished product. For me that's a fair trade, because the knowledge stays with me.

Jarvis and ESIS – what's the difference?

People often ask me why I have two systems. ESIS is my company's operating system: it reads along in the team groups, writes daily reports, flags leads and holds the company memory together. It works for the company. Jarvis is my personal clone. It works for me and with me, on whatever I'm doing right now.

You could say ESIS is the office and Jarvis is the right hand. Both are built on the same basic idea – AI as an employee, not a toy – but they solve different problems.

Why "beta" is the honest label

I call Jarvis beta on purpose. It makes mistakes. Sometimes it understands a task differently from how I meant it. Sometimes a result is too polished, and you can tell there's no experience behind it. Anyone who claims an AI clone is finished and flawless is selling something.

What I learned from that comes down to three rules:

• Clear instructions beat clever models. The more precisely I describe what I want, the better the result. Most bad results were bad instructions.
• Checking stays a human job. Jarvis builds, I review. Especially for anything that goes out the door, there's no "good enough".
• The conversation stays with a human. Jarvis prepares and follows up, but it doesn't sell. I explain why in Sales: emotion beats systems.

Where a clone should stop

A clone can do a lot, but it shouldn't do everything. It shouldn't speak on behalf of people without them knowing. It shouldn't make decisions that involve money, contracts or trust. And it shouldn't pretend to be human. I draw these lines not because the technology couldn't cross them, but because I have to live with the result.

Who a clone of your own makes sense for

A clone of your own pays off for people who have lots of recurring tasks and are willing to describe them properly. It doesn't pay off for someone hoping an AI will tell them what to do. The clone amplifies whatever is already there – clarity as much as chaos.

That's exactly where the Clone Looping System came from: the idea that anyone can build their own clone and rent it out to local businesses. If you'd rather have a finished agent built and maintained for your company, NextCore is the right place.

Jarvis isn't a show project to me. It's proof that a 22-year-old without a computer science degree can build himself the right hand he imagined as a kid. More about me and my path is on About Mario De Luca.