Commerce
E-commerce AI recommendation system
Realtime personalization across millions of SKUs with guardrails for stock and margin.
Problem
A legacy rules engine could not adapt to seasonal demand; merchandising needed explainable bundles without manual curation for every campaign.
Solution
We shipped a two-tower retrieval plus reranker with business constraints, nightly batch retrains, and merchant overrides through feature flags.
Process
How we moved from intake to production scale.
- 1
Data audit
Joined clickstream, stock, and margin signals into a feature contract.
- 2
Modeling
Two-tower retrieval + reranker with business constraints in the loss path.
- 3
Platform
Online inference on GPU, feature flags for merchant overrides, CDN shells.
- 4
Operations
Nightly retrains, monitoring on NDCG and guardrail breaches.
Before vs after
Positioning snapshot—pair with metrics below for diligence.
| Before | After |
|---|---|
| Static rules & spreadsheets | Constraint-aware ML ranker |
| Slow seasonal pivots | Nightly retrains + merchant flags |
| Opaque discounting | Margin-safe bundles with explainability |
Before / after (illustrative)
Architecture (flow)
Storefront → Rec API → feature store → models → warehouse feedback
Clickstream → feature store → online inference (GPU) → CDN-cached shells → feedback loop into nightly trainer.
Tech stack
Results
Higher conversion on recommendation surfaces while controlling discount leakage.
+0%
Rec surface conversion
−0%
Discount depth
↓ major
Campaign build time
- +19% lift on surfaces powered by the new ranker
- Fewer manual campaign builds for merchandising
- Explainable overrides for risky or low-stock SKUs
+19% conversion on recommendation surfaces · −12% discount depth
Screenshots
Illustrative product and analytics surfaces—swap for client-branded assets when sharing externally.
Related case studies
Get a similar solution
Share constraints—we'll mirror architecture, stack, and rollout patterns.