I ship experiences that move conversion, not just features.

Started as a one-tap quick-filter surface placed directly in the listing flow and proved out through a controlled split test on the highest-traffic categories, so the read would be fast and decisive. The validated concept then graduated into a permanent, self-tuning capability, scaled across hundreds of PLPs.

Clear conversion and revenue-per-session lift in test, exposed to 500K+ sessions, then shipped and engaged with by 5M+ sessions to date. What began as a single experiment is now a durable surface that keeps the most-used filters What began as a single experiment is now a durable surface that keeps the most-used filters in the browsing journey instead of buried a tap away.

To make the surface self-tuning rather than hand-curated, I worked cross-functionally with IT to build a data flow that takes standard site-filter engagement data, pipes it into a Google Cloud project, and feeds it back into Algolia as a dedicated index. That index dynamically populates each PLP's most-engaged standard filters as quick filters, so the surface adapts over time to what shoppers actually use, at scale, without manual upkeep.

Owned and monitored a controlled rollout of the agent at scale, synthesizing the experience data for both internal leadership and Google as the external partner. I went deep on the conversational sessions themselves, reading real exchanges to understand what customers came to the agent to do, and where it delivered.

The signal was clear: customers who engaged converted and added to bag at materially higher rates, confirming strong receptivity to the experience. Initial entry-point visibility landed at the industry benchmark for a net-new surface: enough to validate demand, and enough to tell us the next unlock was discoverability, which set up Phase 2: contextually aware prompts placed where intent already lives.

Phase 2 moved the agent from a single entry point to contextually aware PDP prompts, surfacing the assistant at the moments a shopper is already weighing a product, comparing options, or looking for an answer. Given the conversion and add-to-bag lift among users who open it, the program is now scaling into new global elements and surfaces, expanding where the agent meets the customer rather than waiting for the customer to find it.

A/B tested an always-visible search field in the mobile nav across both brands, with two variants. I read results per-brand and per-variant rather than in aggregate, because aggregate would have masked a real divergence.

One brand showed a statistically significant unique-conversion lift, a clear ship. The second brand showed an add-to-bag decline at high significance, which I would not ignore for the sake of a clean launch. Variant 2 showed no lift on either brand.

Shipped the winner, killed variant 2, and held the second brand pending a session-replay investigation into why the add-to-bag metric moved. Holding to diagnose is a decision, not indecision.

Designed a model where engineering owns the actual rendering on critical surfaces and the personalization platform operates only as a server-side allocation and measurement relay, passing results to analytics. Documented it as a PRD with a bow-tie process diagram so engineering, marketing, and analytics shared one source of truth.

Removes the third-party tool from the transaction-critical render path while preserving the ability to allocate and measure tests. Marketers keep their experimentation surface, engineering keeps control of checkout stability.

Rolling the model out MVP-first, proving it on a contained surface before generalizing. The principle: experimentation infrastructure should never be a single point of failure on the path to revenue.