data scientist · luxembourg

I break numbers before I trust them.

I'm the analyst who finds the flaw in the number everyone already trusted. Now I'm taking data science to the problems analytics can't crack alone.

view workdownload CV
LinkedIn ↗GitHub ↗
Bus-delay risk by hour of day, peaking in the morning and evening rush.From my project: predicting bus delays in Luxembourg →

01 / selected work

Two projects. Lead with the data-science proof, back it with the business result.

Machine learning · end-to-end

Predicting bus delays in Luxembourg

Flags ~83% of real delays as a simple daily risk list.

problem

Operations teams usually learn about bus delays only after they have spread across the network. There was no early signal for which trips were likely to run late.

what i did

I collected about 7.4 million stop events myself through a public-transport API, over roughly eight weeks in the summer of 2025, then built one clean, reproducible dataset (SQL to CSV) for both the analysis and the model. I ran EDA across hour, weekday, route, stop, and operator, and engineered 13 leakage-safe features that use only what is known before the bus arrives. I compared a logistic-regression baseline against tree models, scored on F1 (delays are only about 36% of events, so accuracy would mislead), and tuned the alert threshold to a real operations trade-off.

result

A Random Forest risk score that catches about 83% of real delays at the chosen threshold (around 0.46), packaged as a green, yellow, and red alert list with a rollout that runs from shadow to pilot to scale. It also comes with a clear account of where it can't be trusted yet: no real-time traffic or GPS signals, and only about eight weeks of data.

  • Python
  • pandas
  • scikit-learn
  • SQL

Data-science program final project · Dec 2025 · data self-collected via API

view the full project ↗view repo ↗
Business intelligence · cost

Finding a 15% cost leak at a manufacturer

−15% operating expenses.

problem

Rebotec Brasil is a Brazilian manufacturer with a factory and several distribution centers, selling through distributors, retailers, and exports. The company could not see where operating spend was leaking across its sites.

what i did

Built the Power BI dashboards leadership used to track sales, margins and KPIs across commercial and administrative areas; used SQL to extract and clean large datasets; automated recurring financial reports with Excel and Python (pandas); designed and tracked the strategic indicators behind a cost-reduction push; and acted as the primary data partner to senior management.

result

A 15% reduction in operating expenses, and a standing role as the person leadership called when the numbers had to be right.

  • Power BI
  • SQL
  • Excel
  • Python (pandas)

Rebotec Brasil · Business Analyst · 2016-2025

02 / how i work

The same four steps, whether it’s a dashboard or a model.

Whether the output is a KPI dashboard or an ML model, the method doesn't change.

01

Frame

I start from the decision, not the data: who needs it, what good looks like, and what counts as success, before I touch a single row.

02

Build

I get the data myself when I have to, whether that means an API or SQL, then turn it into one clean dataset that everything else runs on.

03

Break

Then I try to break it: baselines, leakage checks, and the metric that won't flatter me. A number survives only if it holds up here.

04

Ship

I turn the result into a decision a non-technical room can act on, like a threshold, a risk list, or a recommendation. Then I say plainly where it can't be trusted yet.

03 / about

Ten years asking what happened. Now I'm asking what happens next.

For nine years I was the person Rebotec Brasil (a Brazilian manufacturer that sells through distributors, retailers, and exports) called when the numbers didn't sit right. I built the dashboards leadership ran the business on, and once found 15% of operating costs hiding in plain sight. Now, in Luxembourg, I'm training as a data scientist, learning to predict what I used to only report.

Brazilian, based in Luxembourg, working in three languages.

  • python
  • sql
  • power bi
  • tableau
  • looker studio
  • pandas
  • scikit-learn
  • excel / vba

04 / capabilities

What I do. Grouped, not rated out of five.

Each of these shows up in the work above, not just on a list.

Analysis

  • Exploratory analysis on messy, real-world data
  • Statistics & probability
  • KPI design and cost analysis
  • A/B testing
  • Feature engineering (leakage-aware)

Machine learning

  • Classification & regression
  • Tree models (Random Forest, Gradient Boosting) vs. interpretable baselines
  • Honest evaluation: F1, precision/recall, ROC-AUC on imbalanced data
  • Threshold tuning to a real decision, not just a metric
  • Knowing where a model can't be trusted, and saying so.

Delivery & communication

  • Dashboards leadership actually uses (Power BI, Tableau)
  • Explaining a result so a non-technical room makes the call
  • Turning a model into something usable: a threshold, a risk list, a recommendation
  • Reproducible workflows and clear documentation
  • Automating recurring reports so they stop eating time

05 / contact

I'm looking for a data science or analytics role in Luxembourg. If that's you, get in touch.

emailLinkedInGitHubdownload CV

luxembourg · pt / en / fr