How I Won the “Mostly AI” Synthetic Data Challenge

A winning solution for the "Mostly AI" synthetic data challenge was achieved by focusing on post-processing rather than model ensembling. The core strategy involved oversampling from a single generative model to create a large pool of candidate data points. This pool was then refined using a multi-step pipeline including Iterative Proportional Fitting (IPF) to match statistical properties, greedy trimming to remove the worst-fitting samples, and iterative refinement to swap in better candidates. For the sequential data challenge, this approach was adapted to first optimize for sequence coherence before applying statistical refinement. Performance optimizations using `numba` and sparse matrices were critical to executing the computationally expensive process within the time limits.