ML-Driven Commerce Platform
- Client
- Samsung Card
- Date
- 2022–2023
- Role
- Operations Strategy + Implementation PMO
- Team
- 50+ people · Consulting, Design, Engineering & Architecture
Context
Samsung Card ran two closed e-commerce platforms: a shopping mall for Samsung Card holders and a benefits mall for employees of Samsung, Hanwha, and affiliated companies. Closed malls — access-restricted, points-funded, curated assortment — generate loyalty, but have a structural growth constraint: you can't grow the audience.
As open marketplaces like Coupang expanded, the closed malls were losing share fast. They couldn't compete on price or logistics. To survive, they needed to make the restriction itself feel valuable — which meant better merchandise, better experience, and lower operational cost.
The Problem
The problem was two layers deep.
First, the two malls served completely different customers but ran on separate systems. The same supplier registered the same product twice, managed inventory in two places, and ran two settlement processes. Onboarding to this closed mall cost suppliers twice the effort of onboarding to an open marketplace. Merchandising competitiveness was impossible when the supply side was this inefficient.
Second, the customer-facing experience was frozen at 2011. More than twenty top-level menu items, a mix of permanent and temporary storefronts, and a login button sitting in the app's prime real estate. No personalization. No continuity between purchases. In a service where 92% of orders came through mobile, the mobile experience was the oldest thing in the building.
What I Decided
1. Merge the backend, keep the frontends separate.
Interviewing suppliers and Samsung Card's operations team made the answer clear. The supplier pain was doing identical work twice across parallel systems — product registration, inventory, settlement. That had to merge. But the customers were different. Benefits mall users spend big with welfare points — they buy premium goods in fewer, larger transactions. Shopping mall users spend smaller amounts on a card, and they price-compare against Coupang. Channel, product, and customer profiles were fundamentally different; a single frontend would satisfy neither. One backend, two frontends. That boundary defined the project.
2. Choose full rebuild over incremental improvement, planning for ten years.
The benefits mall was built in 2011. Patching wasn't viable. Together with Samsung Card, we decided on a ground-up rebuild scoped for the next decade. I wrote new operational policies from scratch across eight domains — membership, product, merchandising, marketing, order and delivery, returns and exchanges, customer service, and shared infrastructure — then ran the build as PMO against those policies.
3. Redesign the interface, introduce ML personalization, and build the measurement layer together.
Cut the top-level navigation from twenty-plus items to seven permanent storefronts. Created a new 'PICK' menu consolidating temporary promotions and personalized events into a single surface. Made the full menu accessible in one click. On top of that structure, we layered ML. We analyzed the distinct Channel-Product-Customer profiles of each mall and built models for menu and category ranking, personalized product recommendations, and targeted event notifications by customer segment. The core customer base was late-30s to late-40s (52.73%), which shaped the content operations strategy. Event storefronts became marketing inventory — modular, redeployable by situation rather than hard-coded. Everything was built responsive, mobile-first. Personalization only works if you can measure it. Samsung Card's closed malls had no analytics infrastructure — no funnel metrics, no cohort tracking, no way to tell whether a recommendation led to a purchase or just a click. As part of the rebuild, we stood up a KPI framework connecting web analytics data (traffic, search volume, duration) with sales and CRM data (conversion, repeat purchase rate, customer lifetime value), structured around three business objectives: customer acquisition, repurchase, and revenue growth. That framework made the ML models accountable — without it, personalization would have been a feature with no feedback loop.
What Didn't Work
We didn't define what a good recommendation meant before building the recommendation model. Under schedule pressure, the decision was to ship the model and evaluate later — but post-launch, without an agreed definition of quality (click-through? conversion? repeat purchase?), the performance conversation had no anchor. The definition should have come first.
The Outcome
Operating costs dropped roughly 27%. Supplier process time was cut by approximately 44% as the parallel workflow collapsed into a single structure. The lower onboarding barrier opened the door to broader merchandise assortment — the thing the business needed most.
What I'd Do Differently
I'd push harder on the business model question at the strategy stage.
This project focused on operational efficiency and experience improvement, and both were necessary. But making a closed mall run better doesn't answer why it should exist when Coupang keeps growing. The unique value of a closed mall — the closure itself — could have been leveraged more aggressively: community built on exclusivity, limited merchandise runs, deeper integration with corporate benefits programs. I spent the project fixing the system and didn't press the business model question far enough.