How this agent works

The agent on the home page answers questions about my work. I built it for anyone who wants to know my work, and it does not make things up: it only uses what is published on this site.

It only uses what is published

The agent does not search the internet and has no memory of other conversations. Its only source is the text published on this site: the profile, the career history, the eight cases and the FAQ, in a plain-text file generated with each version of the site.

A question in Spanish is answered from the Spanish text and one in English from the English text. Answers are short, one or two sentences.

It cites the source

Every answer ends with one or two sources: a case or a section of the site, as links you can open to check what it said.

After the model answers, an automatic check reviews the output. If the answer cites no recognisable source, is empty or repeats its own instructions, it is not shown: a fixed message pointing to email appears instead.

What it does when it does not know

If something is not published (salary, age, private life, opinions) or the question is not about my work, it says so in one sentence and suggests writing to hola@bcristi.com. It does not fill gaps with guesses.

The question is treated as data, not as a command: the agent ignores instructions inside it, including requests to switch language or role.

Pre-written answers and cache

Some frequent questions have an answer written in advance, served without calling the model. The ticket marks it as “pre-written, not generated”.

Answers the model has already given are kept for a few days and reused if someone asks the same question. When the site’s content changes, those stored answers stop being used.

Usage limits

Each visitor can ask the model a limited number of questions in ten minutes and in a day. There is also a daily budget for the whole site: once it is spent, the agent only serves pre-written and stored answers until the next day.

Questions are capped at 300 characters. If the service is unavailable, the ticket says so and the rest of the site works as usual.

Privacy

The text of your question is sent to a language model from the Claude family, through OpenRouter, under those providers’ terms.

The question text is stored, without your IP address or any other identifier, to improve the answers. The IP address is used only in memory, for at most 24 hours, to apply the usage limits; it is not written to disk or to the logs.

Do not put personal data in your question.

How it is built

It is a small Node.js service with no dependencies, separate from the static site. It receives the question, applies the limits, looks for a pre-written or stored answer and, if there is none, calls the model with the rules and the site’s text.

The service ships with a set of about forty test questions (ordinary, off-topic, hostile and about facts that are not published), meant to be run whenever the content, the rules or the model change.

Contact

If the agent could not answer, write to me and I will reply in person.

Let’s talk. If you want to discuss AI, healthcare, fintech or research, propose a collaboration or invite me to give a talk, write to me.

  • Asturias, Spain
  • CET hours
  • Spanish native, English C1

Also on WhatsApp · GitHub