One owned recommender, replacing a five-vendor stack.
A global retailer ran personalization across five vendors with diverging cohorts and runaway cost. We consolidated into one owned, edge-deployed system — and moved AOV by 38%.
Problem, strategy, implementation.
Five overlapping personalization vendors produced inconsistent shopper experiences, diverging A/B winners, and unit economics that broke at scale. Engineering owned none of the models, so iteration cycles were quarterly — not daily.
Replace the vendor stack with a single owned two-tower retrieval + ranker, deployed at the edge for sub-50ms latency. Unify the feature store across web, app and email. Make the team self-sufficient on model evaluation and rollout.
Five Ptimz engineers across modeling, platform and infra. Edge-deployed inference across 14 regions, with warm fallbacks for cold geographies. Full ownership transfer with documented eval, monitoring and rollout protocols.
Kickoff to compounding outcomes.
Discovery
Vendor decomposition, traffic shape, feature inventory, eval baseline.
Modeling
Two-tower retrieval, ranker, calibration; offline + online eval harness.
Platform
Feature store, model serving, observability, A/B framework.
Edge rollout
14-region deployment, vendor cutover, A/B graduation.
Handover
Team enablement, on-call rotation, retrain pipeline ownership.
Outcomes signed off by the client's CFO.
"We replaced a five-vendor stack with one Ptimz system. Order value is up double digits, engineering owns the model, and our infra bill went the right direction for the first time in years."
What we actually built with.
Related case studies.
What could we ship for you?
A 30-minute session with the engineers who would lead your deployment. No slides, no sales — a working conversation.