Personalization Engine Data Flow

Signals and catalog through two-stage candidate generation, ranking, and compliant next-best-action delivery

Personalization Engine Data Flow Signals and catalog through two-stage candidate generation, ranking, and compliant next-best-action delivery 01 / Sources 02 / Features 03 / Generate 04 / Rank + gate 05 / Deliver Customer Signals · behavior + profile · 01 / Sources · events Customer Signals behavior + profile events Product Catalog · offers + terms · 01 / Sources · reference Product Catalog offers + terms reference Feature Store · shared vectors · 02 / Features · point-in-time Feature Store shared vectors point-in-time Candidate Generation · cheap recall · 03 / Generate · top-K Candidate Generation cheap recall top-K Ranking Model · expected value · 04 / Rank + gate · scored Ranking Model expected value scored Eligibility Gate · suitability + fair-lending · 04 / Rank + gate · hard filter Eligibility Gate suitability + fair-lending hard filter Next-Best-Action · explainable offer · 05 / Deliver · decision Next-Best-Action explainable offer decision Delivery Channel · app / email · 05 / Deliver · customer Delivery Channel app / email customer Holdout + Lift · control group · 05 / Deliver · measurement Holdout + Lift control group measurement behavior + profile no protected attrs product attributes reference feature vectors point-in-time top candidates top-K scored offers ranked compliant offers hard filter chosen action explainable outcomes measured Legend primary data policy / PII async batch data store

Two-Stage Path

  • • Recall cheaply narrows the catalog to a small set
  • • Ranking spends compute only on surviving candidates
  • • Labels name data assets, not generic API verbs

Compliance Boundary

  • • Eligibility and suitability are hard filters, not scores
  • • No protected attributes enter features or ranking
  • • Every surfaced offer carries an explanation

Measured Lift

  • • A holdout group never sees personalized offers
  • • Lift is the treated-minus-control difference
  • • Outcomes feed back as future training labels