Portfolio

Launches & systems

A narrative account of products built, decisions made, and lessons learned — across personalization, experimentation, and AI-powered decisioning at scale.

01Amazon2018–2020
PersonalizationAI ProductPlatform JudgmentExperimentation

Amazon Homepage Personalization

Turning the world's most visited retail page into a personal storefront for every customer.

Context

Amazon's homepage was one of the highest-traffic surfaces on the internet — seen by hundreds of millions of customers daily. It had evolved through years of ad-hoc editorial decisions and business unit negotiations, resulting in a page that felt familiar but deeply generic.

Problem

Despite Amazon's vast knowledge of customer behavior and purchase history, the homepage delivered largely the same experience to everyone. Category deals, trending items, and editorial widgets were scheduled weeks in advance, optimized for aggregate clicks rather than individual relevance. Customers with distinct needs — parents of toddlers, home renovation enthusiasts, frequent tech buyers — saw the same surface.

What changed

We redesigned the homepage's ranking and assembly system to be customer-aware at every layer. Dynamic widget selection, slot-level personalization, and real-time intent signals replaced static editorial scheduling. The page became a responsive surface — reflecting not just who you were, but what you were currently thinking about.

My role

Led product strategy and cross-functional coordination across ML science, engineering, design, and business unit stakeholders. Defined the personalization framework, drove the ranking model requirements, and owned the experimentation roadmap. Presented to senior leadership to secure investment and alignment across competing org priorities.

Operating principles

  • ◈Customer relevance over business unit representation
  • ◈Ranking decisions should be explainable, not opaque
  • ◈Personalization is a system, not a feature
  • ◈Every slot is a testable hypothesis

Impact

Significant improvement in customer engagement and downstream conversion. The system scaled to support real-time personalization across customer cohorts — establishing a framework that influenced how Amazon thought about dynamic surface assembly beyond the homepage.

What I learned

“Personalization at scale is as much a coordination problem as a technical one. The hardest part wasn't the ML — it was aligning dozens of stakeholders who each had legitimate claims on homepage real estate, and building a framework they trusted more than the old manual system.”

02Amazon2020–2022
ExperimentationPlatform JudgmentAI Product

Amazon Cart & Experimentation Systems

Building the infrastructure that lets Amazon test at the speed of its ambition.

Context

Amazon's cart and checkout experience is one of the most consequential product surfaces in e-commerce — and one of the most technically complex. The existing experimentation infrastructure had grown organically and was showing its limits: slow test cycles, inconsistent measurement, and experiment interactions that made causal attribution difficult.

Problem

Teams wanted to move faster, but the experimentation platform created bottlenecks. Setting up a new test required significant engineering overhead. Results were hard to interpret due to network effects and overlapping experiments. Leaders lacked confidence in the reliability of decisions made from experiment data.

What changed

We rebuilt the experimentation system from the ground up with a focus on velocity, measurement integrity, and self-service configuration. Introduced stratified assignment, improved variance reduction techniques, and built a decision framework that teams could use without deep statistics expertise. Simultaneously, we made structural changes to the cart experience that improved clarity and reduced decision friction at the point of commitment.

My role

Owned the product vision for cart UX improvements and experimentation platform evolution. Worked closely with applied scientists on measurement methodology. Led the cross-org process change required to move from 'any team can run any test' to a governed, high-confidence experimentation culture.

Operating principles

  • ◈Experimentation quality matters more than experiment quantity
  • ◈Speed comes from removing friction, not from skipping rigor
  • ◈Decision-makers need to trust the system before they'll trust the results
  • ◈Measurement infrastructure is product infrastructure

Impact

Materially reduced time-to-launch for new cart experiments. Improved measurement consistency and stakeholder confidence in test results. Cart UX changes reduced checkout abandonment and increased purchase completion rates.

What I learned

“The best experimentation platforms are invisible to the people using them. If scientists and engineers are spending time on plumbing, you've already failed. The goal is to make the rigorous path also the easy path.”

03Amazon2021
ExperimentationLeadershipPlatform Judgment

A/B Testing Bar-Raising

Raising the quality bar on how one of the world's largest tech companies makes decisions.

Context

At Amazon, experimentation was deeply embedded in the culture — but the culture had outpaced the methodology. Teams ran hundreds of A/B tests simultaneously. The bar for what constituted a valid, decision-worthy result had become inconsistent across organizations.

Problem

Statistical errors — underpowered tests, peeking, multiple comparisons without correction, survivorship bias in reporting — were common. Leaders were making shipping decisions based on results that didn't hold up to scrutiny. Simultaneously, there was social pressure to show 'wins', which created incentives to present favorable-looking results rather than rigorous ones.

What changed

Developed and socialized a set of experimentation standards that became part of how product decisions were reviewed. Created a lightweight decision framework — not a bureaucratic checklist, but a set of questions every team should be able to answer before shipping based on a test. Ran education sessions and embedded this thinking into leadership reviews.

My role

Designed the framework and led the socialization across product and science leadership. Worked with senior applied scientists to get the methodology right, and with VPs to get the organizational buy-in needed to make it stick.

Operating principles

  • ◈Rigor is not the enemy of speed — sloppy decisions are
  • ◈Incentives shape what gets measured and how it's reported
  • ◈Good standards need champions, not just documentation
  • ◈The question is never 'did we win?' — it's 'what did we learn?'

Impact

Improved decision quality in product reviews. Reduced the frequency of questionable test results being used to justify shipping decisions. Established a shared language for discussing experiment quality that persisted beyond the initial rollout.

What I learned

“Changing how an organization makes decisions is a culture change, not a process change. Technical correctness isn't enough — you have to make the right way feel like the normal way.”

04PayPal2022–2023
AI ProductPersonalizationPlatform Judgment

PayPal Payment Ready

Redesigning PayPal's core decisioning to know when a customer is ready — before they ask.

Context

PayPal's core value proposition is making payments simple. But behind every transaction is a complex decisioning stack — credit eligibility, fraud risk, customer lifetime value, funding source optimization — that historically operated as siloed, sequential processes rather than a unified intelligence layer.

Problem

The payment experience was not customer-aware in the way it needed to be. Customers encountered friction — additional verification steps, suboptimal funding source recommendations, delayed approvals — that could have been anticipated and resolved upstream. The system was optimized for risk management, not for customer readiness.

What changed

Designed and launched 'Payment Ready' — a decisioning framework that assessed customer readiness holistically before the transaction moment. Combined risk signals, identity confidence, account health indicators, and behavioral patterns into a unified readiness score that downstream products could act on. The goal: surface payment confidence earlier, remove friction at checkout, and reduce abandonment at the moment of commitment.

My role

Led product strategy for the decisioning framework. Defined the customer readiness model requirements with data science, aligned risk and compliance stakeholders on the framework boundaries, and drove the product roadmap from concept to initial launch.

Operating principles

  • ◈Decisioning should serve customers, not just protect the business
  • ◈Friction reduction is a risk management strategy, not opposed to it
  • ◈Readiness is a state to be anticipated, not just evaluated reactively
  • ◈Trust signals compound — build them deliberately over time

Impact

Improved checkout completion rates for targeted customer segments. Reduced unnecessary verification step triggers. Established a shared decisioning vocabulary across product, risk, and engineering that enabled faster iteration on downstream payment experiences.

What I learned

“The most powerful AI product decisions happen before the user experience begins. If you've waited until the customer is at checkout to decide whether they're trusted, you've already missed the window.”

05Reflection2024
AI ProductLeadershipPlatform Judgment

What I Believe Now About AI Product Leadership

Lessons from building decisioning systems that actually ship.

Context

After a decade of building ML-powered products at companies operating at massive scale, I've developed a set of convictions about what makes AI product leadership effective — and what separates systems that transform businesses from ones that generate demo decks.

Problem

Most AI product efforts fail not because of bad models but because of bad product thinking. Teams optimize for model performance rather than decision quality. They conflate AI capability with AI usefulness. They build in isolation and wonder why adoption stalls. The organizational and epistemological challenges are underestimated relative to the technical ones.

What changed

My thinking has shifted from 'what can the model do?' to 'what decision are we trying to improve, and for whom?' The best AI products I've built or studied share a common architecture: they start with a human decision that matters, identify where intelligence can improve it, instrument the loop carefully, and then iterate based on real outcomes — not model metrics.

My role

Articulating this perspective is its own form of leadership. Writing it down, socializing it with teams, and using it as a filter for prioritization decisions.

Operating principles

  • ◈Start with the decision, not the model
  • ◈The most important metric is rarely the one you can measure directly
  • ◈AI products need theory of mind — who is the human in the loop and what do they need?
  • ◈Durable value comes from AI systems that improve with use, not ones that degrade with scale
  • ◈Responsible AI is a product requirement, not a legal disclaimer

Impact

These principles have shaped how I approach product strategy, team culture, and stakeholder alignment in every AI initiative I've led since.

What I learned

“The leaders who build lasting AI products are not the ones who understand the technology best. They're the ones who understand the human system the technology is embedded in — and who can hold complexity without collapsing it into oversimplification.”