On October 24, 2015, GruPy-SP (with InteropMIX) heard Fabio Franco Costa talk about data science without a slide promising “in three months you will be a scientist.” His thesis fits in one sentence: data science is a method for thinking through problems, not a stack. The video is on the LinuxPro channel. Below is the talk's message — and what still holds true in 2026.
Source: GruPy-SP | Fabio Franco Costa — Data Science in Real Life (26 min). That same evening: Bruno Rocha — Flask and MongoDB.
Who spoke, and where
Fabio Franco Costa, then at CI&T — a software company from Campinas that was already a Brazilian multinational, with around two thousand employees and work mentioned for FIFA and Coca-Cola FEMSA. He presents himself as someone who translates data science for sales and the team: he is not the guy who “knows the answer,” but the one who sets up the test. His Twitter at the time: @costa_2_0. Meetup: GruPy-SP.
The title plays on his own story: data science “in real life” — including the part where he describes using the same method outside the office. It is not a pandas course. It is the craft.
The first principle: you do not know the answer
The barrier, Fabio says, is not the library. It is the attitude of the “doctor who has an answer for everything.” In academia, you are trained to tackle a problem without knowing the function: formulate a hypothesis, measure, fail, formulate again. In business, the opposite is the default — opinion, rank, “we have always done it this way.”.
Data science starts when someone admits: I do not know. Then come the data (almost every company already has tables; an empty starting point is rare), and intuition becomes a question, not a conclusion. He cites the kind of question businesses were already asking in 2015: can you tell whether a customer is pregnant just from her shopping cart? The Target anecdote (USA) was the meme of the time. The point is not American retail — it is turning a hunch into a testable hypothesis.
Hypothesis, then experiment
Having a theory proves nothing. Only the experiment decides. Fabio uses the black swan in the Popper/Taleb sense: “every swan is white” falls apart with a single black swan caught on film. Until you test it, you have a story.
The personal analogy in the talk is an online dating funnel treated as a pipeline: you do not know the DNA of the answer; you build hypotheses (what gets a response, what does not); experiment for a year; measure; discover a side effect (the dinner at home that was not in the sample). The message for the data team is the same: an unexpected result is not a notebook failure — it is the world.
The third step, which he roughly calls walking: if there are no data, you can still design the test. Without a test, statistics cannot save you.
Statistics are not decoration
The more tests you run, the greater the chance of a false positive. Fabio mentions the confidence level of 5%: across dozens of experiments, some “winners” are noise. Without statistics, you are not discussing things with the decision-maker — you are pushing a chart. The scientist advises; the commander decides. Swapping those roles is the shortcut to a vanity dashboard.
The trio he advocates: statistics + business knowledge + research method. Tools alone (the “monkey test”) do not get to the problem. Domain knowledge alone without testing becomes trained guesswork. Testing alone without domain knowledge optimizes the wrong metric.
The simple scoring example
At the end, he points to a classic pediatric test — what the audience recognizes as Apgar (Virginia Apgar, 1952): a few objective questions about the newborn, a score from 0 to 10, rather than just the on-duty doctor's “clinical eye.” The literature of the time already showed that a simple checklist, applied immediately, reduces first-day mortality compared with a specialist's unstructured impression. Fabio's message: a humble model + data in the right place often performs better than a well-trained human on autopilot.
It is not “AI replaces the doctor.” It is this: the scientific method fits any initiative, and the best place to apply it is where there is no legacy yet — a new product, a new process, an internal startup.
A data-driven organization
His “personal crusade” in 2015: helping companies make some decisions with humility and evidence, not just intuition. Data-driven does not mean BI on the wall. It means reducing the number of choices made in the dark.
CI&T was launching a product based on data science / machine learning in the following cycle. The name on the box matters less than the habit: a written hypothesis, a designed experiment, statistics in the middle, domain knowledge at the end.
What to do with this on Linux in 2026
The setting was Python — GruPy. The local path remains the same, with newer versions:
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -U pip pandas scikit-learn matplotlib
Tutorial and venv: Python from Beginner to Professional. The language overview: The 7 main programming languages.
Pandas and sklearn still handle everyday tabular work. What has changed is the number of ready-made models (and people who skip testing). The 2015 talk is the antidote: I do not know → hypothesis → experiment → a number that stands up to discussion.
Wrap-up
A GruPy Saturday, a guy from CI&T, a method academia already teaches and business forgets. Watch the video. Then open a dataset of your own and write the hypothesis before the chart.
GruPy-SP: meetup.com/Grupy-SP · channel: youtube.com/linuxpro.