Retailers rarely have a promotion problem or a fraud problem in isolation. They have a checkout decisioning problem where margin, conversion, and risk collide in the same session. The data already exists across commerce, OMS, fraud, and CDP. The issue is deciding in one pass before the shopper abandons.
Retailers rarely have a promotion problem or a fraud problem in isolation. They have a checkout decisioning problem where margin, conversion, and risk collide in the same session. The data already exists across commerce, OMS, fraud, and CDP. The issue is deciding in one pass before the shopper abandons.
This is built for digital commerce, payments, fraud, and growth teams. The visible demo stays customer-safe. The script is where the presenter carries the detail, the stakes, and the ask.
The data is already there: Commerce platform, OMS / inventory system, promotions engine, fraud APIs, customer profile store, feature lakehouse, and Kafka checkout events. The issue is not whether the data exists. The issue is whether it can be assembled and acted on in <10 ms while the moment is still live.
That is what this demo shows. Nine stages, one primary decision moment centered on Priya Shah, and a decisioning pipeline that turns scattered operational context into the winning action: Free expedited shipping with no discount.
This is the full architecture for what you are about to see. At the top are the systems of record and live signal sources — the commerce platform, order management system, customer data platform, fraud service, inventory and fulfillment APIs, and Kafka clickstream events. None of these go away. They stay exactly where they are.
The ingest layer has two jobs. RDI handles change data capture and operational sync from the core repositories — always-fresh context with no custom pipeline code. Redis FeatureForm handles the feature pipeline from the analytical and streaming systems into the context layer, with full train-serve parity. Two tools, two roles, one unified ingest layer.
The context layer is the operational working set. Hot cart state and live session signals stay in Redis RAM for sub-millisecond access. Larger purchase history, browse embeddings, and warm behavioral context sit in Redis Flex. Redis Context Retriever connects those stores to the decision engine as the semantic access layer — it assembles the Shopper 360 — cart state, purchase history, and checkout policy — and exposes it as structured MCP tools the decision engine can call directly.
The decision engine is where eligibility rules, ML ranking, and policy arbitration come together. The output channels are where the shopper sees the result. And the learning loop makes every accepted or rejected action improve the next one.
This is the architecture. Now let me show you what happens when the live customer moment actually starts.
This is Priya Shah. Loyalty member with a two hundred eighty-six dollar cart of mixed apparel and home goods, and she just hit checkout. This moment has to resolve before the checkout page finishes rendering.
This is not an edge case. This is the repeatable decision moment Northfield Commerce handles hundreds of thousands of times a day. The system has to be fast enough to stage the right action before Priya sees the page — because what she sees next determines whether she completes the order or abandons it.
If the system waits too long, it falls back to a generic discount. If it acts on partial context, it surfaces a ten percent basket coupon that erodes margin when Priya would have converted on free shipping alone. If it decides in time with full context — purchase history, live fraud score, inventory state, and promotion policy all assembled together — it captures the checkout at better margin.
We have one live moment to recognize Priya Shah correctly and act before the old process falls back to something generic.
Northfield Commerce keeps everything you see at the top of this architecture. The commerce platform, order management system, loyalty platform, fraud and payment APIs, and Kafka clickstream events all stay exactly where they are. Redis is not the new system of record. Redis Iris is the context engine — the operational serving layer that makes those existing systems act together in the live checkout window.
The ingest layer has two jobs. RDI handles change data capture from the commerce platform, OMS, and customer data platform — always-fresh context sync with no custom pipeline code required. Redis FeatureForm handles the feature pipeline from the analytical systems and streaming events into the online feature store, with full train-serve parity. Two tools, clear separation of concerns, one unified ingest layer.
The result is a working set that is always current. Not a nightly batch. Not a stale snapshot. Milliseconds behind the source.
Redis does not replace the existing stack. RDI and Featureform make that stack operational in the live decision window.
This is the heart of the decisioning stack. The left panel is Priya Shah's durable profile — customer value band, relationship tenure, prior interaction pattern, eligibility state, policy constraints, and promotion frequency history. The right panel is what is happening right now in this session — current intent, live inventory state, capacity availability, risk and compliance check, and surface readiness.
Most systems have one or the other. They can look up a customer record. Or they can capture a live checkout event. The gap is serving both together at request time, inside the latency budget.
Redis Context Retriever is what makes that possible. It assembles the Shopper 360 — cart state, purchase history, and active checkout policy — and surfaces it as structured tools the decision engine can call directly. No fan-out queries, no manual joins across repositories. History without live state is stale. Live state without history is shallow. Redis is the layer that serves both in the same response path.
A profile tells you who the customer is. Context tells you what the business should do next.
You are looking at six features served live from Redis in under a millisecond each — cart value, fraud risk, inventory fit, shipping eligibility, promo elasticity, and customer LTV band. One hundred eighty-six features total across this decision path. P99 lookup latency under twelve milliseconds.
The point is not the feature names themselves. The point is that these are the same features used to train the model, served online at decision time with the same definitions and the same logic. That is train-serve parity. Most teams can train a model. The hard part is serving the right features fast enough in production without drift between the notebook and the live application.
Redis FeatureForm on Redis closes that gap. These features are specifically why the system chooses Free expedited shipping with no discount instead of defaulting to the ten percent basket coupon or surfacing a buy-now-pay-later prompt that does not fit this session.
Your model is only as good as the features you can serve in milliseconds, not the features you can describe in a slide deck.
The winner is Free expedited shipping with no discount, with an NBA score of 0.94. It wins because it fits this exact moment — Priya is a gold-tier loyalty member with a two hundred eighty-six dollar cart, low fraud risk, strong inventory fit, and high promo elasticity. Shipping is the incentive that converts her without surrendering basket margin to a blanket coupon.
Ten percent basket coupon scores 0.79. That is the default path when the system has incomplete context. It converts, but at a lower contribution margin than shipping when the shopper is loyalty-qualified and inventory-ready.
Buy-now-pay-later prompt is suppressed. Fraud risk is low but basket mix and session signals do not warrant it. A model operating on partial context might have surfaced it. The full picture removes it before it reaches the decision engine.
Redis Search is what powers the similarity matching in this ranking step. Vector search is not a separate product you bolt on — it is a query type that Redis Search handles natively, the same way it handles full-text and numeric filtering. It is just another data type Redis can search at sub-millisecond speed.
This is not content ranking. This is checkout decisioning — arbitrating across fraud posture, margin, fulfillment fit, and promotion policy in one low-latency response.
We are not surfacing random recommendations. We are ranking the actions the business already cares about and choosing the one that fits this moment best.
These numbers are direct results of the architecture. Decision latency of 9.8 milliseconds means the checkout action is staged before the page finishes rendering. A 9 point checkout conversion lift means fewer abandoned carts — especially in the sessions where the right incentive is all that was needed to complete the purchase. And 6.70 dollars of margin protection per order means the platform is not reaching for the highest discount to close every transaction.
The value is not this single checkout. It is what happens when this decision gets repeated across millions of sessions — every cart, every loyalty tier, every fulfillment window where the system is choosing between a margin-preserving action and a generic fallback. That is where the math compounds.
That is also why the next step is a pilot, not a deeper technical evaluation. The question is not whether Redis is fast. The question is what one checkout flow looks like when the decisioning layer runs on live context instead of static promotion rules.
The math is not the single transaction in front of us. It is what happens when this decision gets repeated across the full book of business.
Same shopper. Same checkout page. Same moment. The left side shows what happens without the context layer — partial profile, delayed retrieval, limited live signals. The system surfaces a ten percent basket coupon because that is the safest generic option available.
On the right, the same checkout page opens with the right action already staged. Free expedited shipping with no discount — best conversion lift with minimal margin loss. Checkout conversion lift of 9 points is the visible result.
The product is not the UI. The UI is identical on both sides. The product is the decision layer underneath it — the one that assembled cart state, fraud score, inventory availability, and promotion policy before the page finished loading.
Same surface. Same moment. Different decision layer. That is the product.
This is the same architecture you saw at the start. Every tier looks the same. What is different now is that you have seen what each one contributed to the outcome.
Three takeaways. First, this is not a science project. This is a practical reference architecture that Northfield Commerce can operate today. Second, it is additive — the commerce platform, OMS, CDP, and fraud service stay exactly where they are. Redis sits in the operational path so those systems can act together. Third, this is a business story first. Higher checkout conversion, 6.70 dollars of margin protection per order, and fewer defaulted discounts are the reasons to do it — not the platform architecture.
The next step is a focused working session to map this against your actual environment. We scope one checkout flow, one basket segment, and one pilot that runs Redis-powered checkout decisioning alongside your current promotion logic. That is a clean comparison with a real KPI before you commit to broader rollout.