Practice 02
Spreadsheets in, analysed reports out, on schedule — with every figure traceable back to the row that produced it. Because a number nobody can verify is worse than no number at all.
Scope agreed in writing before work begins.
The problem
The weekly report is the same shape every week. Someone exports the data, cleans it, pastes it into a template, writes the same three observations, converts it to PDF and emails it. Multiply that by the number of people doing it and the number of weeks in a year, and you are paying a salary for work a system should do.
The trap is automating it badly: an AI that produces confident summaries nobody can check. The output has to be traceable to the source, or you have replaced slow work with fast risk.
Scope
Stated before you ask, so the first call is about your problem rather than about what we do or do not cover.
How it runs
Not the whole department. One report, one workflow, one inbox — the one people complain about by name.
People, hours per week, fully loaded cost. That number is the budget and the success criterion at the same time.
Ingestion, rules, AI analysis where it adds value, and validation that flags anything the model is not confident about instead of hiding it.
For a few cycles the manual version and the automated one run side by side and the outputs are compared. You switch when the numbers match, not when the demo looks good.
Track record
Built an agentic AI assistant that analyses large databases, performs precise calculations and turns the result into VP-level executive reporting — the kind of output that gets read in a leadership meeting rather than filed.
Certified in AI Agents with RAG and LangChain, plus Python (Pontificia Universidad Católica de Chile) and Databases for Data Scientists (University of Colorado Boulder). Working automation stack: Python, n8n, RAG pipelines.
Author of a modular AI operating framework built around skills, playbooks, decision trees and reviewers — the discipline that stops AI-assisted delivery from being brilliant one week and unusable the next.
Questions
That is the risk the design has to answer. Figures come from your data through code, not from the model's memory; the model writes the narrative around numbers it did not invent, and anything it is unsure about is flagged for a human rather than smoothed over.
That is decided with you before anything is built, and it is written into the scope: which model provider, what leaves your infrastructure, what is retained and for how long. If nothing may leave, the design changes accordingly.
An MVP typically starts around USD 4,500 and the exact figure depends on how many systems it has to touch. The assessment that precedes it starts around USD 1,500 and you keep the roadmap either way.
Systems built here are meant to be edited: documented, versioned and handed to your team. Small changes should not require calling us back.
Start here
Describe the problem in a few lines. You get a written reply with a first read on it, whether or not there is an engagement in it. The message is sent from this page — no email client, no third-party form service, no trackers.