Jose Antonio Licon · Pittsburgh, PA

One person.
Five very different projects.
One engine.
Shipping daily.

I build apps, games, and platforms — powered by an AI development system I built myself.

See the workThe engine →

Xanadu isn't an AI assistant. It's an autonomous development loop.

Not “Tony uses AI to help him code” — but a tiered, self-correcting, rule-learning execution system that compresses two weeks of work into two days of wall-clock time.

It runs a BAD_PATTERNS immune system that learns from every mistake. A velocity feedback loop that shapes what gets built next. A cost cascade that starts with a local 7B model, escalates to 32B, and only reaches Claude or Gemini when genuinely stuck — so it's fast, cheap, and doesn't waste frontier compute on problems it can already solve.

The apps it ships aren't just products. They're the benchmarks. Galaxian at 535 completed tasks is a live proof of concept for the system itself.

What you're watching in the video above isn't a demo of AI-assisted development. Most people describing that mean Copilot or Cursor. This is categorically different — an unattended system with a planner, an executor, a validator, error classification, and a rule promotion loop.

That's Xanadu.

Projects

Retro Car Radio

Live

Internet radio with a classic car preset interface. Old school vibes, a world of streaming.

iOSAndroidFlutter

Galaxican

Coming soon

A space strategy game where stars produce resources, fleets capture territory, and one more turn becomes five.

iOSAndroidFlutter/Flame

Magic Task Hat

In development

Personal productivity powered by Agile principles. Your backlog, your sprints, your rules.

FlutterFirebaseGCP

Estate Wise

In development

Property management for 8 real buildings and 29 real units. Built because nothing else fit — and it actually works.

Next.jsFirebaseGCP

Podomus

Moonshot

Free podcast hosting on GCP free tiers. Start your show. Zero cost, zero excuses.

GCPFirebase

The engine

Xanadu

An autonomous development loop: local LLMs in a tiered cascade, mechanical error correction, and a planning layer that turns a backlog into running code — unattended.

Claude and Gemini sit at the top, handling escalations and architecture. Everything below runs locally. A 2-week plan compresses to 2–3 days of wall-clock time.

Tasks completed autonomously

535+

First-pass success rate

~70%

LLM tiers in the cascade

4 + Claude

Why I built it

I started where everyone starts — Claude, Gemini, the usual suspects. I was amazed at what was possible. Then I ran out of tokens.

Most people slow down at that point. I bought a maxed-out MacBook Air and started learning local LLMs instead.

What began as a workaround turned into something more interesting: a complete development loop. Product ideation. SWOT analysis. Backlog generation. Code execution across parallel workers. Models that fail over to more capable tiers when a task is too hard. Error patterns that get learned and encoded so they don't burn retries twice.

Now I set it running overnight. By morning, the project is mostly done. When I need more firepower, I spin up RunPod. A week of backlog in an hour, at a cost that doesn't require a VC.

It didn't replace the craft — my background as a full-stack developer and technical product manager is what makes the system work, not what it replaced. I still drive. I still take the wheel when the models hit a wall, or the project needs a pivot. Agile methodology is the backbone of the whole system — not just a buzzword, but the actual structure that keeps five projects moving at once.

This isn't a silver bullet. It's a superpower.