A new AutoML system called VirnyFlow refuses to pick between an accurate model and a fair one; it optimizes for both, plus stability, at the same time.
Most AutoML tools lock in a single objective before the search even starts, usually raw accuracy, and treat fairness or stability as an afterthought. VirnyFlow instead lets a user-defined evaluation protocol, including fairness measured across intersectional groups, guide the entire search: Bayesian optimization of individual pipeline components, bandit-based selection of pipeline structures, and pruning that weighs multiple criteria at once. It runs on Apache Kafka for asynchronous, fault-tolerant execution and logs everything to a database so developers can inspect trade-offs mid-search. In tests on six real-world datasets, it matched or beat established frameworks including auto-sklearn, Alpine Meadow, and FLAML under identical resource limits, and scaled to 128 workers across four nodes on datasets with 2.6 million records, running up to seven times faster than the best single-node alternative.
That scaling number is less interesting than the philosophy shift. Rather than handing back one best model, VirnyFlow lets data scientists interrogate the trade-offs themselves and adjust the search to fit a specific deployment, something the researchers tested with a clinical case study on distribution shift and an IRB-approved user study. That is a meaningfully different pitch than typical AutoML, which just picks a winner, and it matters more as regulators start asking how fairness got measured, not just whether it did.
Whether that control survives contact with a messy production pipeline, well outside six benchmark datasets and a university IRB, is still an open question.