Medellin.co

Traditional content websites need constant manual maintenance, and a city guide is close to the worst case. Restaurants close. Hours change. New businesses open. Neighbourhoods evolve. Events move. Prices drift. Recommendations quietly go stale.

When Medellin.co ran on WordPress, most of that work was manual. A single business-directory update could mean:

1. Searching Google for the business.
2. Checking whether it was still open.
3. Verifying the address.
4. Checking the website.
5. Checking the hours.
6. Reviewing social media.
7. Updating WordPress.
8. Updating images and related information.

Researching and updating one business properly could take about 45 minutes. That does not scale to hundreds of businesses, and it certainly does not scale to thousands.

Medellin.co became version two of the Nova Platform. Rather than relying on a traditional CMS workflow, the site uses AI agents, structured data, external APIs and automated research to manage its own information.

The goal was never to add AI features to a website. It was to make AI part of the website's operating system.

Automated research

The platform researches information from configured sources and uses it to help maintain and expand the site. Instead of searching manually every time something needs updating, agents handle most of the initial research, and can flag or update content depending on the workflow.

That moves the human role away from repetitive research and toward verification, strategy and editorial judgement.

Self-healing content

One of the ideas behind Nova is that a website should be able to detect problems with its own information. The system looks for content that has become outdated, incomplete or inconsistent, and opens a workflow to investigate it.

Rather than waiting for someone to stumble across a stale page, the site takes part in monitoring itself. That became the foundation of what I call a self-healing website.

Automated business directory

The directory is the clearest example of what the new architecture bought.

It is integrated with Google Places, so the platform can retrieve and verify business name, address, location, trading status, hours, reviews, ratings, contact details and website information automatically.

Verification that took about 45 minutes by hand is now largely automatic. I spend my time on the parts that need actual local knowledge or editorial judgement instead.

Kathe, the AI concierge

Travel websites get very specific questions. Someone might ask: where should I stay if I want to be close to restaurants but not in the busiest part of El Poblado?

A traditional site makes that visitor assemble the answer from four articles. Kathe lets them just ask.

Kathe uses retrieval-augmented generation. Rather than leaning on a model's general knowledge, it retrieves from Medellin.co's own articles, guides, business information, neighbourhood data and internally researched content before answering. The website itself becomes the source of what the assistant knows.

The experience changes shape entirely, from search → article → article → article to question → research → answer.

From website to knowledge platform

The most important shift was in how I think about websites at all.

Traditional sites are built for humans navigating pages. But AI agents are now another kind of visitor: search engines, assistants, personal agents and automated research systems all retrieve directly from websites.

So I started designing Medellin.co for two audiences, people who browse it, and AI systems that retrieve from it. That decision shaped how content, business data and internal knowledge are organised throughout the platform.

For years websites have followed the same model. Someone creates content. Someone publishes it. Someone maintains it. Someone eventually notices it went out of date.

Agentic systems break that model. A website can actively research, monitor, organise and maintain its own information.

The biggest lesson was that AI does not need unlimited autonomy to be useful. Giving each agent a clearly defined responsibility, and keeping a human where judgement or verification matters, works better than handing over the whole job.

I also learned how much structured information matters once AI is part of the interface. If the underlying data is organised properly, the same knowledge can drive web pages, search, AI assistants, business directories, content generation, recommendations and whatever agents come next. The website stops being a collection of pages and becomes a knowledge system.

Medellin.co started as a city guide. It became a test of a much larger question: what happens when the website itself becomes an intelligent system?

Instead of needing a person to constantly research, update and organise information, the website takes part in the work. It is one of the projects that shaped how I build agentic AI websites, and eventually the Nova Platform itself.