Nine storefronts, one component library, no per-brand forks.

Every brand in the group needs the same furniture: a product card, a cart, a checkout, a review block, a search surface. What differs is the identity on top - the palette, the typography, the tone.

Built brand by brand, that produces nine of every component, drifting apart with every release. A fix lands in one storefront and not the other eight. An accessibility improvement is made once and forgotten eight times. The per-product override becomes the normal way to get anything done, and the overrides are where the inconsistency accumulates.

The work was to replace that with one shared foundation that all nine storefronts consume, themed rather than forked.

I architected a multi-brand React component library with dynamic theming, so a single component serves all nine product lines and takes its identity from the brand consuming it rather than from a copy of itself.

It follows atomic design, is built on Tailwind CSS, is documented in Storybook so a developer on any brand team can see what exists before writing a new one, and is held to accessibility compliance rather than having it retrofitted.

Around the library sits the rest of the shared foundation - the supporting packages every storefront needs, versioned and released so brand teams can adopt changes deliberately instead of being surprised by them. The result is plug and play: the per-product manual overrides are gone across all nine.

Alongside that I optimised Core Web Vitals across the storefronts, which moved page load, SEO and the ranking signals that follow from both.

A shared library only holds if what people add to it matches what is already there. That is a hard enough problem with humans; with AI assistance it gets harder, because a model with no context will happily invent a ninth button.

So I authored agent skills that encode the architectural patterns, coding standards and domain knowledge of the codebase, so generated work arrives context-aware and production-shaped rather than plausible-looking.

Underneath, I built a knowledge dependency graph over the shared source and the guardrails that read from it, so an agent can find what already exists instead of generating a near-duplicate. That is also where the efficiency came from: 70% less token usage, because retrieving the right context beats re-reading everything.

I led and mentored the developers working on the shared architecture, with unit and end-to-end coverage over it.

The guardrails were the highest-leverage thing built and the last thing started. Retrieval and consistency enforcement should have come before broad agent use, not after the first wave of near-duplicate components made the case for them.

Nine consumers also means nine adoption curves. I would invest earlier in making version upgrades trivial for a brand team, because a shared library that is awkward to adopt quietly becomes a forked one.