Bogdan Gaza

Before DatologyAI, Gaza was CTO and cofounder of Moonsense. His employer biography also identifies more than a decade of infrastructure engineering and management at Amazon and Twitter. At DatologyAI, he works with a team studying how the choice and processing of training examples affects a model’s quality and the computation needed to train it.[1]

DatologyAI’s text-curation technical report credits Gaza for leadership and advising. The team combines model-based filtering, embedding-based filtering and synthetic data in a production pipeline. It starts with RedPajama V1 and compares models trained on the curated output with models trained on several public datasets, keeping model families and training budgets explicit. This separates the effect of the data from simply making the model larger.[2]

The report evaluates the resulting models on language-model benchmarks and asks three related questions: whether better data improves accuracy, reduces training needed for a target score, or lets a smaller model reach stronger performance. Its experiments extend to 2.7-billion-parameter models and 180-billion-token budgets. Those bounds define what the reported results establish, rather than proving the same gains for every model or workload.[2]