He Zhu has spent eight years designing analytics and workflow systems for users who need evidence before they act.

LLM product design. Mika is the AI agent he designed inside Logica, Kubit’s self-serve analytics platform. It stays grounded in customer data, shows confidence indicators, and keeps a human in the loop. He owned the prompt patterns, the retrieval context, the review states, and the evidence users want before they trust an answer.

0→1 ownership. Five years as Kubit’s solo designer, plus coaching one junior designer and two interns. He picked the constraints, wrote the principles, and shipped the product. The result was a $4M ARR product with 85% logo retention, built from a blank page. At Timo he is doing that blank-page work again, owning strategy, IA, and the design system through to ship.

Data-dense B2B SaaS. Self-serve analytics, deep admin consoles, enterprise role-based permissions, multi-tenant workspaces, no-code formula editors. His tokenized design system with 40+ components supported a 50-person team and cut front-end implementation time roughly 40%.

LLM-aware UX. This is the work he finds most exciting right now. A model that hallucinates, sounds certain when it is wrong, and drifts over time is the kind of messy, high-stakes design problem he likes most. Show the evidence behind every answer, signal low confidence and offer a fallback, and put evaluation where everyone can see it, not just the ML engineers.

His bar is simple. If a user can’t see why the model answered the way it did, the model didn’t answer. It asserted. At Kubit, Mika shows the metric definition, dataset window, and chart-construction logic next to every reply. The formula editor shows a live preview of the syntax tree as you type, because the model shouldn’t be the only thing that knows what your formula means. Aegis Clinician applies the same rule in a clinical setting: AI Scribe must move through Draft → Review → Accept → Sign before any note locks.

None of that needs AI hype. It needs the discipline any high-trust system asks for.

0→1, blank-page work. He’s most useful when the product question hasn’t been asked yet. He runs discovery with founders and engineering leads, names the constraints, and sketches the IA against how the operation actually runs. Then he ships a working prototype that answers the “what is this” question with evidence instead of slides. The job is to cut ambiguity enough for the team to build.

1→n, scale-up phases. Once a product earns a second customer, the shared system becomes part of the product. This is where he builds the design system, the tokens, the components, and the governance, and where he writes the principles and patterns the rest of the team can extend without him. At Kubit the system supported a 50-person team across his tenure. At Timo he’s building that same substrate from scratch.

When AI is in the loop. That means grounding answers in real data, designing fallbacks for when the model is wrong, keeping evaluation visible, and handing off cleanly to engineering through AI-assisted prototyping (Figma Make, Cursor, Claude Code, Claude Design, Lovable). The work has the same shape as any other systems work. The difference is that the failure modes come from model behavior, not UI state alone. The Glance case study documents exactly where that loop helps and where it stops.