AI agents

A concrete job, not a decorative chat

I design and implement an agent for a real task in your operation: answer from your information, query a system and hand the case to a person when needed.

What this is

An AI agent, in this package, is a worker with one concrete task. It answers from your information, can query a system, and keeps a record of what it did. When a case leaves what it knows, it hands the case to a person. It is not a generic chat stuck on a website as decoration. The scope is one agent, one source — documents, a spreadsheet, a database or an API you already have — and a clear handoff. Several agents, several channels, or training your own model sit outside and are discussed first.

Experience

I am technical product owner of ARIADNA, a B2B SaaS for learning and human capital. The work is turning a need into scope, data contracts, acceptance criteria and tests on real data, with a person in the loop when automation must not decide alone. At Falabella I managed delivery of AI products and the corporate technology portfolio. This package uses that same discipline, hands-on in TypeScript, React and Next.js, not as a generic demo.

How I work

I use two ways of ordering the work. They are not creeds. Each one fits a different moment.

Waterfall

First we write the whole path and what sits outside it. Then we build in that order. It fits when the scope can already be named: a six-screen site, an integration between two systems, an Odoo start with three modules. You know what you will receive before we begin.

Scrum

The work is split into short weeks. Each week something usable is shown, and we decide what comes next. It fits when the detail is still being found: an agent's tone, the question that actually reaches the bot, the priority in a backlog.

The published package is the frame: scope, timing, and what is quoted separately. That is waterfall. Inside it there is a weekly review: we look, we correct, we carry on. That is Scrum. I hold a Scrum Master certificate and Scrum Foundations, plus a Green Belt for when the problem is operational rather than fashionable.

Who it is for

Teams that already lose hours answering the same thing, or that want to try AI on documents, a spreadsheet or an API that already exists.

What the package includes

Scope from real cases, not a generic demo
Instructions, limits and the agent's tone
A connection to one source: documents, a spreadsheet, a database or an API
Tools to query and, where it fits, to record
A place to test, correct and see what it answered
Handover of how to run it and what it must not touch

Quoted separately

Several agents or several channels at once
Training your own model
High token volumes, which are paid to the provider