Building a multi-team server-side experimentation platform

Problem

Estée Lauder's product teams each ran experimentation in isolation — some on client-side tools, others not at all, none on a shared decision-making framework. The brands the company runs span dozens of properties: own-brand commerce sites, marketing surfaces, app experiences. A client-side-only approach couldn't reach the server-rendered and headless surfaces, and shipping experiments through engineering for each one was the bottleneck blocking velocity.

Approach

I took over the Experimentation team and made two structural changes. First, I restaffed and repositioned the team around platform thinking instead of project execution — fewer one-off implementations, more reusable infrastructure other teams could consume. Second, I directed the development of a server-side experimentation platform built in Elixir with gRPC. The platform exposes a small, opinionated API so any team — Shopify, custom storefronts, internal tools — can register an experiment, get a deterministic variant assignment, and emit exposure events to the shared analytics pipeline.

In parallel, I revamped the client-side experimentation program to focus on throughput. Faster QA loops, lighter-weight templates, fewer hand-offs — the goal was to move experiment ideas from product brief to results in a fraction of the previous cycle time.

During the Shopify migration I led the effort to leverage AI to build integrations that let business users self-serve experiments directly from the Shopify UI — translating a designer's intent into a valid experiment definition without requiring engineering tickets.

Outcome

• A reusable platform unblocking experimentation for teams that previously had none
• Client-side throughput up significantly; decision-making accelerated across the org
• Business users running their own experiments on Shopify without engineering involvement

Technical highlights

• Elixir + gRPC for the platform: low-latency variant assignment, supervisable, easy to scale horizontally
• Deterministic bucketing keyed on a stable user identifier so server and client always agree on the variant
• AI-assisted authoring layered onto Shopify so business users describe an experiment in natural language and get a structured definition back

What I led, not just shipped

Repositioning a team is a leadership job, not an engineering one. The technical work mattered, but the bigger lever was deciding what the team wouldn't do anymore.