Notes on data engineering, AI and building things end to end.
A random train/test split lets a time series model peek at the future — which inflates its accuracy in a way that never holds up in production. Here’s how proper time series cross-validation actually works.
Read more →
Labeling the four boxes of a confusion matrix is the easy part. Here’s how to actually read one — what each cell costs, and how to tell if a model is genuinely good enough — with real project examples.
Read more →
New users and new items both break collaborative filtering the same way — no history to work with. Here’s why the cold start problem happens and the techniques that actually help, with real project examples.
Read more →
Most data portfolios run once on a laptop and stop there. Here’s why CI/CD — automated tests, reproducible environments, version tracking — matters for data pipelines just as much as application code, with real examples.
Read more →
SHAP plots look precise, which makes them easy to over-read. Here are the three most common SHAP misinterpretations — and what these values actually tell you, with a real project example.
Read more →
DuckDB skips the infrastructure bill a cloud warehouse demands — but it’s not a universal replacement. Here’s when dbt + DuckDB genuinely wins, and where it stops making sense, with real project examples.
Read more →
Accuracy lies on imbalanced datasets. Here’s what to measure instead, which techniques actually help, and how to pick the right precision/recall tradeoff — with real examples.
Read more →
Most data science and data engineering portfolios fail on the same five mistakes — misleading accuracy metrics, untested pipelines, black-box models, and unverifiable results. Here’s what to check instead, with real examples.
Read more →
A notebook that prints an accuracy score is a script. Here’s what actually separates it from a data product — serving layers, version tracking, usable interfaces, and visible reasoning — with real examples.
Read more →