Lifted Conversion Rate by 34% Without Touching the Core Catalog System
Project Overview ML-Based Recommendation Engine
Most e-commerce platforms don’t have a “we need AI” problem. They have a “our personalization is a rules engine wearing an AI costume” problem — static “customers also bought” logic, hardcoded category cross-sells, and merchandising rules that haven’t changed since launch, all quietly capping revenue per visitor.
This is the ceiling MoraStack was brought in to break.
A mid-sized e-commerce platform carrying roughly 60,000 SKUs across twelve categories had outgrown its rule-based recommendation logic. Recommendations were the same for every visitor regardless of browsing behavior, purchase history, or session intent. The product and growth teams suspected this was suppressing both conversion and average order value, but had no way to test that hypothesis without an entirely new system.
MoraStack was engaged to design and build a machine learning–based recommendation engine — not as a storefront redesign, but as a backend inference layer that plugged into the existing catalog and checkout systems without requiring either to be rebuilt.
Client Requirements – ML-Based Recommendation Engine
The client came in with three hard requirements and one underlying constraint that shaped every architectural decision.
Hard Requirements
- Replace static, rule-based product recommendations with behavior-driven, personalized recommendations at the individual visitor level
- Support multiple placement contexts — homepage, product detail page, cart, and post-purchase email — each requiring different recommendation logic
- Provide a feedback loop so recommendation quality improves continuously from live interaction data, not periodic manual retraining
Underlying Constraint
The existing catalog management system and checkout flow were stable, heavily integrated, and non-negotiable for replacement. The client’s data warehouse (Snowflake) was also fixed infrastructure. Any system MoraStack built had to read from these systems and serve inference results back into the storefront without becoming a second source of truth for product or order data.
This constraint ruled out most off-the-shelf recommendation SaaS tools, which typically require full catalog migration into their own data model, and made a custom inference layer sitting alongside the existing stack the only viable path.
Problems Identified
Before any model work began, MoraStack conducted a two-week audit of catalog structure, historical interaction data, and existing recommendation logic across engineering and growth teams.
Static Rule-Based Logic
Recommendations were generated from a fixed set of category and co-purchase rules last updated over a year prior. No mechanism existed to surface emerging product affinities or seasonal shifts.
No Behavioral Signal Capture
Clickstream and session data were logged for analytics but never structured or fed back into recommendation logic. Years of browsing and purchase signal existed but sat unused.
One-Size-Fits-All Placement
The same recommendation logic powered every placement — homepage, PDP, cart, and email — despite each context implying a different visitor intent (discovery vs. cross-sell vs. win-back).
No Experimentation Framework
There was no infrastructure to A/B test recommendation strategies, meaning the team had no way to validate whether changes to logic actually moved conversion or AOV.
Cold-Start Blind Spot
New visitors and newly listed products received generic, non-personalized treatment with no fallback strategy beyond category bestsellers.
Engineering Approach
MoraStack’s principle here is direct: build the inference system, not a single model. A point fix — one collaborative-filtering model bolted onto the homepage/ solves one placement and leaves the rest of the funnel running on static rules. System engineering means building a serving layer that supports multiple models, multiple placements, and continuous retraining, with the catalog and checkout systems untouched.
Data Layer
A unified feature store aggregates clickstream, purchase history, catalog metadata, and session context from Snowflake into a structure designed specifically for model training and real-time inference — without altering the source systems.
Modeling Layer
A hybrid approach combining collaborative filtering for established products and users with content-based filtering for cold-start cases, blended through a ranking layer tuned per placement context.
Serving Layer
A low-latency inference API serves recommendations to each placement in real time, with placement-specific business rules (inventory availability, margin thresholds, merchandising overrides) applied as a final pass after model ranking.
Feedback & Retraining Layer
Click and conversion events from served recommendations are logged back into the feature store, feeding a scheduled retraining pipeline so model quality compounds over time rather than degrading.
Experimentation Layer
Built-in A/B testing infrastructure allows the growth team to run controlled comparisons between model variants and the legacy rules engine, with statistical significance tracking surfaced on a dashboard.
Methodology – ML-Based Recommendation Engine
Phase 1 — Data and Behavioral Audit (Weeks 1–2)
Full inventory of available interaction data, catalog metadata quality, and existing recommendation logic, conducted with engineering and growth leads. This audit directly shaped the feature store schema.
Phase 2 — Architecture and Model Design (Weeks 3–4)
MoraStack delivered a full technical specification covering the feature store schema, model architecture for each placement, the inference API contract, and the retraining pipeline design. Signed off by the client’s engineering lead before implementation began.
Phase 3 — Model Development and Offline Evaluation (Weeks 5–9)
Models trained and evaluated offline against historical data using held-out validation sets, with precision@k and recall@k benchmarked against the legacy rules engine’s historical performance before any live exposure.
Phase 4 — Shadow Mode Deployment (Weeks 10–11)
The inference layer ran in parallel with the live rules engine, generating recommendations without serving them, to validate latency, throughput, and ranking quality under real production traffic.
Phase 5 — Phased Live Rollout with A/B Testing (Weeks 12–14)
Live traffic split between the legacy rules engine and the ML system, beginning with the homepage placement before expanding to PDP, cart, and email. MoraStack remained on-call through the full rollout window.
Solution Implemented
The final system is ML-Based Recommendation Engine inference service deployed alongside the client’s existing infrastructure, reading from the catalog and order systems without modifying either.
Unified Feature Store
Behavioral, transactional, and catalog data are consolidated into a single structure purpose-built for both training and real-time lookup, refreshed continuously from Snowflake.
Hybrid Recommendation Models
Collaborative filtering handles established products and repeat visitors; content-based filtering handles new products and first-time visitors, with a ranking layer blending both per placement.
Real-Time Inference API
Recommendations are generated and returned within the latency budget required for homepage and PDP rendering, with business-rule filtering applied post-ranking.
Continuous Feedback Loop
Every impression and click is logged back into the feature store, feeding a scheduled retraining cycle so the system improves from live behavior rather than going stale.
Experimentation Dashboard
Growth and product teams can launch, monitor, and conclude A/B tests on recommendation strategy without engineering involvement per test.
Results for ML-Based Recommendation Engine
The system reached full production rollout across all four placements at the end of week 14. Results measured over the following 60 days:
Conversion Rate Lifted by 34%
On sessions exposed to ML-driven recommendations versus the legacy rules engine, measured across homepage and PDP placements.
Average Order Value Increased by 19%
Driven primarily by improved cart and post-purchase email cross-sell relevance.
Click-Through Rate on Recommendations Up Over 2.1x
Compared to the static “customers also bought” logic it replaced.
Cold-Start Coverage Reached 92%
Of new visitors and newly listed products received personalized or content-based recommendations within the first session, versus generic bestseller fallback previously.
Manual Merchandising Overhead Reduced
Category teams no longer manually curated cross-sell rules for the majority of placements, freeing an estimated 10 hours per week previously spent on manual rule maintenance.
Future Projections (12 Months)
Based on current traffic volume and the performance data from the first 60 days, MoraStack projects the following at the 12-month mark:
Compounding Model Quality
Recommendation precision improving further as the feedback loop accumulates a full annual seasonal cycle of behavioral data.
Revenue Attribution Growth
ML-Based Recommendation Engine attributed revenue share of total online revenue continuing to climb as placement coverage expands to additional touchpoints such as search results and abandoned-cart flows.
Reduced Dependency on Manual Merchandising
Near-complete phase-out of manually maintained cross-sell rules across all placements, with merchandising teams shifting toward strategic curation rather than rule maintenance.
Scalability Without Re-Architecture
Headroom to extend the same inference layer to new product lines or markets without re-building the underlying feature store or serving infrastructure.
Work With MoraStack
If your recommendations haven’t changed since launch, the problem isn’t your catalog — it’s the absence of a system that learns from what your visitors are already telling you.
MoraStack specializes in machine learning infrastructure and inference systems for e-commerce, media, and product-led businesses. We build recommendation and personalization engines that treat behavioral data as a continuously compounding asset — version-controlled, observable, and built to improve with usage, not against it.
Every engagement begins with a data and behavioral audit. We map what signal you already have before we design what you need. No assumptions, no black-box SaaS lock-in, no rip-and-replace.
Engineering as systems, not features.
→ Contact MoraStack to scope your recommendation engine project.
Disclaimer
This case study represents a composite of real project work conducted by MoraStack. Client-identifying details including company name, catalog specifics, and team structure have been anonymized. Quantitative outcomes reflect actual measured results from the described engagement. Projected 12-month figures are estimates based on measured baseline data and current traffic volume; actual results will vary based on operational context. MoraStack makes no guarantee of identical outcomes for future engagements.
FAQs – ML-Based Recommendation Engine
What is an ML-based recommendation engine in the context of software engineering?
An ML-based recommendation engine is an inference layer that uses behavioral, transactional, and catalog data to generate personalized product recommendations in real time — replacing static, rule-based logic with models that improve continuously from live interaction data.
Can a custom recommendation engine integrate with an existing catalog and checkout system?
Yes. MoraStack builds recommendation engines as inference layers that read from existing systems rather than replacing them. The catalog, checkout, and data warehouse stay in place while the recommendation layer consumes and serves data alongside them.
How does a hybrid recommendation model handle the cold-start problem?
Collaborative filtering, which relies on historical interaction data, struggles with new users and new products. Content-based filtering fills that gap by ranking recommendations on product attributes alone, ensuring new visitors and new listings still receive relevant suggestions from day one.
How long does a recommendation engine implementation typically take?
MoraStack’s standard engagement for a mid-complexity catalog runs 12–14 weeks from data audit to full production rollout, including offline model evaluation, shadow-mode validation, and a phased, A/B-tested live cutover.
Why do static, rule-based recommendations stop performing as a catalog grows?
Rule-based logic is manually authored and rarely updated, so it can’t account for shifting product affinities, seasonal demand, or individual visitor behavior. As catalog size and traffic grow, the gap between what static rules suggest and what would actually convert widens, directly suppressing conversion and average order value.