Our latest patent, Predictive Models for Biological Experiments, just published through WIPO. I’m listed as a co-inventor alongside Micah Iriye and Daniel Drobnicki.

The core problem we set out to solve is one that anyone running long-duration scientific experiments knows well: how do you accurately recognize revenue when the timeline, scope, and cost of an experiment are constantly shifting? Traditional approaches peg revenue to a fixed schedule, but scientific work rarely follows a fixed schedule. Supplier performance varies, change orders come in, delays happen.

Our approach uses machine learning models trained on historical experiment data to continuously predict both the total invoice value and the expected duration of an experiment. Revenue is then recognized daily along that projected trajectory rather than on a rigid timeline. As real-world conditions evolve throughout the life of an experiment, the model recalibrates and the revenue curve adjusts accordingly.

The system supports a range of model types including regression models, decision trees, random forests, and neural networks, and works across a broad set of assay types: cell-based, genomics, proteomics, imaging, clinical laboratory, drug discovery, metabolism and pharmacokinetics, and immunology.

This isn’t a theoretical exercise. We built this because we needed it. It’s been running in production as part of our revenue recognition system at Scientist.com, keeping us accurate and compliant even as experiment outcomes shift under our feet.

You can find the full publication on WIPO’s PATENTSCOPE and on Google Patents.