How do you get 25 million people to follow clinically sound recommendations?

A recommendation can be clinically correct and still go nowhere, because whether a person acts on it depends on barriers the clinical logic never sees. Our client, a digital wellness platform serving 25 million members, was running its recommendations on a single one-size-fits-all behavior model: the same assumptions applied to every member, regardless of who they were or where they were in their journey. They asked us to replace it with a system that models each member individually and plugs directly into their recommender.

We built a behavioral engine from scratch: the decision logic that determines how many points an action is worth, what surfaces for a member today, which trigger fires and when, and how all of it adapts as the member changes. Each member is represented as a behavioral phenotype that updates over time. Every candidate action is scored by a personalized cost function on activation energy as well as two other proprietary barriers, and priced by a reward engine where points equal clinical weight times difficulty times sponsor priority. Before anything touched the product, we ran the full engine against synthetic members in a validation harness, so design failures surfaced in an afternoon instead of six months into production.

The engine now powers the platform's recommender in production, replacing the single static model with per-member decision logic. Because the client owns the engine's specification - the phenotypes, the cost function, and the scoring rules - their team extends it independently: new actions, new conditions, and new member segments slot into the same logic without a rebuild.