The best customer experience isn’t one people use — it’s one they can’t live without.

The customer truth

You can’t design an experience you haven’t watched.

There isn’t one journey — there are dozens, and the same person lives several. The first-time visitor. The power user who reorders on autopilot. The shopper comparing five humidifiers. The viewer three episodes into Coupang Play. The one starting a return. Each leaves a trail of signal, and a real CX practice watches all of them — not just the funnel that’s easy to chart.

The cognitive tax is highest in the gaps between the steps: the comparing, the second-guessing, the is-this-the-right-one. Speed is a symptom of getting it right — a new user reaching their first purchase in minutes — not the goal itself. The goal is to find every place a customer is spending effort the experience should be spending for them.

The shopping journey — where the tax lives

1. First launch

   Unknowns. No signal yet to personalize from.

2. Signup & WOW membership

   Commitment before value is felt.

3. Search

   Translating a fuzzy intent into the right query.

4. Product page

   Reading reviews, weighing price vs. quality vs. delivery — the comparing.

5. Cart

   Second-guessing. Did I pick the right one?

6. Checkout

   Friction Coupang has already largely solved.

7. The wait & unboxing

   Rocket Delivery — the won battle.

8. Reorder

   Re-deciding what should already be remembered.

The thesis

Most commerce AI answers questions. The opportunity is to take the work of deciding off the customer.

Amazon bolted a chatbot onto a storefront — a faster search box. That’s not the unlock. The unlock is an experience that anticipates: it reaches the product page already knowing what this customer weighs — reviews, delivery speed, price, brand, quality — and surfaces it first, summarized, half-decided. An AI block inside the page that does the reading, so the customer doesn’t have to.

And when the customer arrives with a goal instead of a query, the assistant carries it the rest of the way. Find a gift for someone you barely know. Pull together everything for a kid’s birthday this weekend. Turn the week’s dinners into a single cart. The customer names the outcome; the experience absorbs the dozens of small decisions underneath it. That is the moment someone says: how did we ever live without this.

Pillar · Anticipation & personalization

Use generative AI to reason. Use classical models to rank. Knowing which is which is the job.

Not every problem needs an LLM, and pretending otherwise is how you blow the budget and lose the customer to the wait. Deciding what to show, and in what order, is a ranking problem — a recommender does it faster, cheaper, and more reliably than any language model. That’s most of the surface area.

Generative earns its cost somewhere narrower: reasoning about a goal, weighing messy human criteria, explaining a trade-off in plain language. Reviews-first for one shopper and delivery-first for another is ranking. The meal-plan assistant is reasoning. The craft is routing each job to the model that does it best — and never paying LLM prices for work a smaller one nails.

Pillar · Measured relentlessly

The AI learns the patterns. People decide what they mean.

Every step is instrumented — not to surveil, but to see what’s actually working. The AI surfaces the patterns; people decide what they mean, because a pattern is never a reason on its own. Experimentation gates the rest: A/B, causal inference, feedback loops.

“Did they buy” is the weakest signal in the set. Did they come back unprompted? Did they wander into a new category? Did they trust the assistant enough to hand it the next decision? That’s the scoreboard that matters.

None of this has to be built from scratch. A self-hosted experimentation platform — GrowthBook, for one — gives you feature flags, A/B assignment, and Bayesian readouts while keeping every event on infrastructure you control. Self-hosted on purpose: the measurement layer has to honor the same privacy line as everything else.

The hard parts

The vision is the easy part. Here’s what’s actually hard.

Put an LLM in the core funnel for tens of millions of customers and it’s a latency-and-cost problem long before it’s a feature. So you cache hard, route to small models by default, and spend the expensive calls only where they change the outcome.

You hold a hard line between reducing decision fatigue and manufacturing it: the assistant serves the customer’s goal, even when that means a smaller cart today. And you localize for real — Korean shopping behavior isn’t American behavior with translated buttons.

And you hold the long view. AI in the funnel invites a parade of short-term asks — investors and customers want the flashy feature now. A Director sees the trade-off the room doesn’t: which of those wins a quarter and quietly costs you trust, latency, or the next year. You ship the ones that compound, and you say no — out loud, with the reasoning — to the ones that don’t. Naming the hard parts is how you prove the vision is built, not dreamed.

Why me

I build this kind of thing. You’re using one right now.

This portfolio runs on a customer-facing LLM I designed and shipped — guardrailed, grounded, evaluated against a 116-question test set, cost-managed with prompt caching, with a privacy classification layer deciding what it will and won’t say. The same discipline this role needs, already in production.

Before that: a checkout and donation flow redesign that lifted completion 35%; a government platform built 0-to-1 that cut a core workflow by 95% and scaled to millions at 100% WCAG; and a habit of anchoring every decision to a real person, not a persona.

Where I’d ramp: operating an LLM at full B2C scale, and deeper fine-tuning — both adjacent to work I’ve already shipped, not new terrain. And Korea isn’t an abstraction to me. It’s personal.