Sensory-First
Our methodology sits within the sensory pairing tradition rather than the molecular pairing hypothesis. We ask "what do ingredients taste like to people?" and "how do chefs and bartenders like to use these flavors?" We believe this produces profiles that are more meaningful, more practical, and more aligned with how flavor decisions are made in real bars and kitchens.
Statistical Rigour
We apply several layers of statistical analysis to ensure our profiles are robust. Bayesian scoring produces confidence intervals for each flavor dimension, reflecting not just how strongly a note is present but how certain we are of that assessment given the volume and consistency of the evidence. We also run hold-out validation, building profiles from independent subsets of data to confirm they converge.
Expert Human Correction
No algorithm fully captures the nuance of flavor. Our Flavour Society members can review and edit any ingredient profile directly, correcting inaccuracies that the computational pipeline may have introduced. Every edit is logged, and patterns in member corrections are fed back into our models, creating a continuous improvement cycle where human expertise refines the algorithm and the algorithm gets smarter over time.
The result is a methodology that combines the scale and consistency of computation with the irreplaceable judgement of experienced palates, always grounded in how flavor is perceived.