Christian Rocco

AI in practice

Real examples of how I use AI at work. The yellow steps are the checks: the places where a rule, a guardrail or a person decides before anything moves on.

How I build with AI

I'm not a developer. I know what needs to exist, I break it into steps, I test what the AI builds, and I take it to the point where people use it.

  1. Decide what's neededA manual task, a missing report, a check nobody runs.
  2. The AI builds itCopilot CLI or Claude Code writes the code, the formula or the flow.
  3. I test and review itDoes it do what was asked? Is the data right? If not, back to the AI.
  4. The team uses itAnd I write down how it works.

Agents that do the program's technical work

At TCS I run a set of AI agents that build DAX measures, Power Automate flows, data pipelines and documentation. Every teammate uses at least two of them.

To make them reusable, I keep a library of 73 skills: written instructions that tell an agent how to do one job well. I use them every day.

The controls around the agents

Agents fail in repeated ways: they make up data, skip rules, and repeat old mistakes. Over 6 months I built a control system around them, one layer at a time. Each layer came from a real failure.

  1. Written instructionsWhat the agent should and shouldn't do.
  2. Rules from real failuresEach time something went wrong, it became a written rule.
  3. Automated checksWhen a rule wasn't enough, it became a check the agent can't skip.
  4. Guardrails that blockRisky actions, like deleting files or pushing changes without approval, are stopped before they run.

All of this runs inside TCS security policies, not around them.

Matching problems with past solutions

Often a problem someone is stuck on was already solved somewhere else. I built an LLM-based model that matches open blockers with cases already solved and proposes a fix. It's in active use.

  1. An open blockerSomething stopping the work.
  2. The model compares itAgainst cases that were already solved.
  3. It proposes a fixAnd estimates its impact.
  4. A person decidesThe model suggests. People choose what to do.

Getting 12 years of sales out of an old ERP

A personal project from 2026. Years of sales history were locked inside old database engines. With AI, I moved the data into a single 2.2 GB SQLite database and built an executive sales report going back to 2014, plus an agent that can answer questions about the data with read-only access.

  1. Legacy ERPOld database engines and files.
  2. One SQLite database2.2 GB, in one place.
  3. Read-only agentIt can look, it can't change anything.
  4. Executive sales reportFrom 2014 on, with notes on what the data can't tell you.