The challenge
The existing recommendation engine optimised for click-through and was quietly training itself toward clickbait — high engagement, poor conversion, and inventory that did not match what was being pushed.
Any replacement had to serve recommendations inside the page render budget, which ruled out calling a large model synchronously on every request.
Approach
01
Separate the slow path from the fast path
Heavy reasoning runs asynchronously to build and refresh user-level preference representations. The request path does a fast lookup and rank, which is how sub-100ms is achievable at all.
02
Reward the outcome, not the click
The RLHF reward signal was built on completed purchases and retained items rather than engagement, which is what corrected the clickbait drift.
03
Join recommendation to inventory
Recommendations are constrained by forecast availability, so the copilot stops promoting what is about to go out of stock.
04
Ship behind a holdout
Rolled out against a holdout group from day one so the conversion and inventory numbers are measured, not asserted.
Architecture
Signals
Real-time behaviour events stream into the feature store alongside catalogue and inventory state.
Learn
RLHF pipeline tunes ranking against purchase and retention outcomes on a scheduled cadence.
Reason
AWS Bedrock models generate the copilot's conversational layer and preference summaries asynchronously.
Serve
Precomputed candidates ranked at request time, staying inside the page render budget.
Forecast
Demand forecasting feeds availability constraints back into the ranking layer.
Results
+14%Conversion
+31%Inventory accuracy
62%Better forecasting
<100msRecommendation latency
A/BMeasured against holdout
What I took from it
Changing the reward from clicks to completed purchases was a one-line change in the pipeline and the single largest driver of the conversion lift. Most of the engineering effort went into making that measurable.
AI Copilot delivered a 14% boost in conversion and 31% inventory accuracy improvement within 90 days of go-live. Every recommendation is grounded in data and commercial reality.
Stack
Models
Orchestration
Cloud
Data