Projects

Case studies on real public data

Each one starts from a real question, uses real public data, and shows the answer, how it was checked, and every line of code.

Model validation

Can we trust an alert model, and the AI that explains its alerts?

A full validation of a machine learning threat model and of the LLM assistant that summarizes its alerts: an honest test on later data, where it fails, the cut-off trade-off, and four tests of the AI.

Known attack types77%New attack types34%
Share of attacks caught at the default cut-off
1.00 → 0.96AUC on a random split vs the later test file
77% vs 34%attacks caught: known types vs new types
10 of 10times the AI said “not in record” instead of inventing an answer
Data: NSL-KDD (Canadian Institute for Cybersecurity, UNB) · Google Gemini 3.1 Flash-Lite
XGBoost vs logistic regressionOut-of-time testFailure analysis by typeCut-off trade-offLLM tests: consistency, groundedness, guardrails
Read the case study →Code
Business analytics

Which leads become customers, and how fast?

A small sales funnel dashboard on real data: win rate by lead source, time to close, rep concentration, and a simple forecast that caught a May shortfall and found half of its cause.

182183April175130May
Customers by lead month: forecast vs actual Forecast Actual
13.0%of 5,998 leads became customers
12 daysmedian from first contact to a signed deal
175 vs 130May forecast vs actual; three reps who stopped closing explain about half
Data: Olist Marketing Funnel (Kaggle, CC BY-NC-SA 4.0): 8,000 leads, 842 won deals
pandasWin rate by sourceTime-to-close curveRun-rate forecastGap analysis
Read the case study →Code