Rockfish is built on 6+ years of peer-reviewed work from Carnegie Mellon on time-series generation, rare-event synthesis, and privacy-preserving generative models — published at venues including AAAI, IMC, and CAIS. The science is why the data holds up under evaluation.
Published at the field's top venues
The reason Rockfish data survives real evaluation comes down to three research pillars — each addressing a place where generic synthetic data falls apart.
Generative models that preserve temporal structure, multivariate correlations, and domain behavior — not just row-level statistics.
Techniques for generating the anomaly spikes, failure cascades, and edge cases that almost never appear in production data but break your models.
Learning the patterns, not the people — so outputs carry the statistical shape of your data without exposing real records.
The foundational work behind the platform — read it yourself.
How to generate realistic, domain-specific evaluation suites that expose failures in AI analytics agents before they reach production — using synthetic time-series and grounded Q&A pairs.
Read the paper →The foundational research establishing that generative models can produce realistic networked time-series — statistically faithful enough to replace or augment production datasets for training and testing.
Read the paper →A technique for synthesizing realistic examples of rare occurrences — the anomaly spikes, failure cascades, and edge cases critical for model robustness.
Read the paper →The science isn't a footnote — it's the engine. Here's how each research pillar became a product you can use.
See how the research translates into evals you can actually ship on.