Trafikanté: A Personal AI Filter Against Information Noise
A practical example of an AI agent that filters sources, evaluates relevance, and sends briefings based on work context.
Information FOMO is often solved by adding more sources. Another newsletter. Another LinkedIn profile. Another Slack channel. Another spreadsheet of tools that “need to be tracked”.
But when working with AI and technology, the problem is usually not missing information. The problem is the lack of a filter. A filter that understands what a person is currently working on, which topics are relevant, and what can be safely ignored.
Trafikanté, Denisa Hrubešová’s personal agent, is built on this principle. In the source material, it is described as a system that reads sources, sorts updates, and sends short briefings to Telegram. Not as a large application. More like a personal information assistant that knows the work context.
When following sources stops being enough
Denisa needs to follow several areas at once: an AI newsletter, a podcast studio, Forendors, and the broader creator economy. As the CEO of a tech company, she wants to know what is happening, but she does not want to spend the entire day manually going through sources and deciding what deserves attention.
That is why Trafikanté does not just produce a list of links. For each item, it explains why it might be relevant. This is an important difference. A normal feed shows what is new. A good filter shows why something should matter to you.
That is where the value lies. Not in following more. More in letting fewer things enter your attention without a reason.
What Trafikanté actually does
The workflow is practical and fairly straightforward. A robot goes through RSS sources, removes items that have already been seen, and sends the rest through Claude to evaluate relevance and produce a short summary. The result arrives in Telegram as a briefing.
The source material mentions six briefings per day on working days. That may sound like a lot, but the point is not to overwhelm the user. The point is that every message has already passed through a filter. Instead of manually reading the internet, the user receives a pre-sorted overview.
The user can then respond to the message in natural language. For example, by asking to expand on a topic or save it into the right context. The important thing is that this is not another dashboard you need to remember to check. The workflow arrives in a place where it is natural for the person to process it.
Why Telegram makes sense
The transcript makes clear that Telegram was not chosen for effect. It was the easiest to implement, and it also separated agent messages from the work environment. Slack is the channel where Denisa has colleagues and things that need attention. Telegram can wait.
That is a practical decision. When designing an AI workflow, people often ask whether the solution is sufficiently “enterprise”. But the channel should also be chosen based on the type of attention the workflow requires. If a briefing does not need to be handled immediately, it should not sit among urgent team messages.
Minimalist technical architecture
Trafikanté is also interesting technically. The source material describes an intentionally minimalist stack: Python, a few libraries, the filesystem as a database, Markdown files for notes, and JSON for state. No large database, no standalone app, no server.
Orchestration runs through GitHub Actions. The Telegram Bot API handles briefing delivery and the reply poller. The Claude API handles summarisation, intent classification, and deeper research.
This is an important point for teams that immediately imagine a large project when they hear “AI agent”. Sometimes, an ordinary folder of files, a few clear rules, and a well-designed flow of information are enough.
Where the risks are
Automated research without review can go in the wrong direction. One participant in the discussion shared an example where her agent, after some time, started recommending or creating links to articles that did not exist. This is not a detail. Trust is the foundation of any information workflow.
That is why Trafikanté is not only about automation. It is also about correction. Denisa described how she iterated the agent at the beginning, marked irrelevant outputs, and added what she was interested in. She selected sources. Some she accepted, others she rejected. This is how a tool becomes a filter.
If the agent does not know what is relevant, it will produce polished summaries of the wrong things.
How to start without building your own agent
Not everyone needs an automated Python setup immediately. The first step can be much simpler.
Choose five sources that are genuinely worth following. Then write down three work topics that are currently relevant to you. Finally, define three topics you will deliberately ignore, even if they are popular around you.
That can become the first prompt for a personal research filter. For example:
I follow these areas, these topics are important to me right now, and I do not want to focus on these. For every new item, tell me why it is relevant, whether it is worth saving, and what question I should ask about it.
Trafikanté shows that a good AI workflow does not need to start with a large platform. It can start with a clear decision about what you allow into your head and what you do not.