Vladimir Dietrich

Automation engineer. Twenty years inside a federal law enforcement agency, removing manual work from systems that could not afford to fail.

Now I want to point that at the hardest problems left: repairing the human body, and building machines that think.

Selected work

2006–07

Courts and police, connected in real time

Brazil's Federal Justice and Federal Police exchanged criminal records by official letter. I conceived and led the replacement end to end — the design, the coordination between institutions, the authorizations, persuading leadership on both sides — from a regional office, far from headquarters in Brasília. The result: a direct, secure connection to the national criminal information system (SINIC): court staff query records in real time and register judicial decisions straight into the national database. A pilot in 2006, more than a year of emails, calls and meetings, then one signature at a time: the agreement was signed in 2007 as a state pilot and later became national, formalized with the Superior Court of Justice, the Federal Justice Council and the Regional Federal Courts.

Read the case →
2008

An indicator that only worked in the US, made to work everywhere

Dr. Alexander Elder's New Highs–New Lows indicator had no working version outside US markets. I built one that pulled daily data and computed it for Brazil, Australia, Canada, China and others. Elder reviewed it personally, endorsed it, and included it in the indicator's official e-book.

Read the case →
2010s

Ten times more cases solved: making invisible work visible

In a public forensic team, I built the case database and a live dashboard that followed every case from the first call to the conviction — every step in between measured, counted against every call that dispatched the team, evidence or not. At the start, only a low single-digit share of cases ended with an identified suspect.

The funnel showed exactly where cases were lost — not where everyone assumed — and justified new equipment at that one step. Team-level numbers helped. Then each person got a bar with their own name on it, and the data showed some of us — me included — had been hiding in the group average. Within about three years the rate was roughly ten times higher. When someone identified a suspect, the dashboard played that person's chosen song for the whole team.

Along the way: seasonality analysis to forecast the week's workload like a weather report, and the discipline of not trusting samples under 30 cases.

People rarely hide from work. They hide from ambiguity.

2025

A blog that publishes itself

I write in a Google Doc. A pipeline reviews the text with AI, generates the cover image, translates it, compresses the assets and publishes. The blog and the project pages you are about to click run on it.

Read the case →
2026

One AI thinks, another one types

An open-source harness that lets Claude Code hand mechanical work to a cheaper agent (Gemini Flash) and get back only a short summary, with the full diff on disk for review. The default model is chosen by a benchmark built from traps: a rename where one identical key must not change, and a refactor with invariants no test covers, checked by a hidden judge. Among models that deliver the same diff, the fastest wins.

See the code →

Seeing the curve early

“Artificial intelligence increasingly has less ‘ifs, loops, and a programmer adding functions and codes’.”

From Job goes away, production stays, July 2020 — two years before ChatGPT — arguing that AI would automate even the programmers. In 2021, Neuragame imagined games played straight through a brain–computer interface. I have been writing about where humans and machines meet since 2018, in the blog.

What I believe

The best automation is invisible: the problem simply stops happening, and nobody has to learn anything new.

I care about repairing and regenerating the body — not about freezing youth. Live each phase fully, even if a phase lasts a million years; what I distrust is a life spent postponing itself to the next million. (Eternal Youth: A Tyranny?)

Background & tools

Languages
TypeScript / Node.js, Python, Apps Script
Cloud
Google Cloud, Firebase, serverless functions, queues
AI
LLM pipelines with Claude and Gemini, local models on GPU, multi-agent delegation
Automation
Playwright, Android, messaging APIs
Education
Electrical Engineering, Escola Politécnica, University of São Paulo (USP)
Spoken
Portuguese (native), English (fluent)

Contact

Brazil, working remotely. hello@vladi.dev