What I learned building ExoSeeker

ExoSeeker started as a hackathon project and ended up teaching me more about feature engineering than any course did.

The Kepler light-curve data is noisy and imbalanced — real planets are rare. The wins came from class weighting, careful validation splits, and an ensemble of a PyTorch MLP with scikit-learn’s HistGradientBoosting rather than chasing a single big model.

90% classification accuracy, Best Use of NASA Data, and a ticket to the global stage.

If I rebuilt it today I’d spend even less time on the model and more on the data pipeline. That’s usually where the accuracy is hiding.

(Placeholder post — replace with your own writing.)

← all transmissions