I turn manual workflows into AI-native systems of action.

Who designs the agentic layer for a workflow-heavy B2B SaaS platform?

That's exactly where I work: AI-native human+agent operating layers that reduce manual work without removing human control.

Recently, I led UX/UI design at Agentic Dream on Strata, an AI-driven B2B SaaS platform for enterprise operations and support.

What does human-in-the-loop design actually look like in production?

It looks like a screen, not a principle. On Repfabric I designed a decision-history screen that gives every major action a visible, auditable trail — so anyone can see what was decided and why, after the fact. That is the human-in-the-loop part: not a confirmation dialog, but a record the team can audit.

The judgment call sits one level up. I sequence AI adoption across three levels — fixed-rule automation, agents making bounded decisions inside a single step, and adaptive loops that weigh options against history, SLAs and permission scopes. On Strata I shipped the middle level into production, where the decision surface was narrow enough to trust, and I designed the third one with audit trails and permission scopes but held it back until the review step could be removed safely.

So in production it is two things: a place where decisions stay visible, and an explicit line about which decisions the system is not allowed to take on its own.

How do you decide when AI should not act?

I sequence AI adoption across three levels. Deciding where to stop is the design work.

On Strata I shipped L2 into production, where the decision surface was narrow enough to trust. I designed L3 — and held it back. The audit trails and permission scopes weren't mature enough to remove human review safely.

Shipping less was the decision that kept a live operation stable.

How do you decide what not to build?

The same discipline shows up in how I scope. On Strata, research surfaced 14 distinct users and 25+ pain points; I scoped the model down to the 3 driving adoption and designed for their control points instead of trying to solve everything at once.

On Repfabric, my research into who actually used the platform versus who bought it brought a 20-role CRM/ERP down to the one persona who drove adoption — a call reached in conversation with the client, weighed against the cost of building for all twenty.

Two projects, the same move. That makes it a method rather than a one-off, and it's usually the decision that makes everything after it shippable.

How do you keep a Figma design system in sync with the code that ships?

I create multi-tier design systems in Figma — the collaborative surface where I explore, design, and iterate with the client — then ship them as code: tokens, components and a Storybook. Once a component is resolved, Storybook becomes the versioned source of truth engineering builds from; Figma stays the surface for the next round of exploration.

The system was structured for direct engineering consumption: semantic variables for color and spacing, annotated component states, and named interaction patterns, carried from Figma into Storybook. Every component had a documented state model aligned with the backend workflow states.

This meant engineering could implement directly from Storybook without translation — reducing back-and-forth and keeping design intent intact through shipping.

It holds because both sides consume the same semantic tokens. It's a contract, not automation: it breaks the moment engineering re-derives a value instead of consuming the token.

What do you actually deliver when you deliver a design system?

Taking a component library and dressing it in a brandbook is the general job. The layer above it is different: no component library ships it on its own. It has to define the states, permissions, provenance, and behavior that let humans and agents work from the same contract.

When I reviewed them in August 2026, IBM Carbon, AWS Cloudscape, Atlassian Rovo and GitLab Pajamas all documented AI patterns. All four were missing the same thing: uncertainty, hallucination risk, the limits of confidence.

You only get this class of defect building a layer that did not exist. They are proof it was built, not installed — and neither surfaced from using the site.

Who can design AND build a production frontend without an engineer?

I researched, designed, built, deployed and tested sebastianbalderas.com end to end on my own — Next.js, design tokens mapped straight to code, and the conversational layer you're using right now. No engineer, no team. That's my own site, not a platform with a team behind it. The claim isn't that I replace engineering — it's that I can build what I design, well enough to know what I'm asking for.