Selected systems we've put into operation.

Problem. Multiple product teams shipping in the same brand had drifted into separate visual languages. Cross-team features required negotiation on basic primitives every time.

System. A token-first design system with three-layer naming (primitive → semantic → component), a Figma library connected to code, and a contribution protocol. Editorial rules included alongside the components — same kind of governance applied to copy.

Outcome. Screen build time fell substantially, and the system has been maintained by the in-house team without our involvement.

Problem. The team was triaging a high regular inbound volume by hand — most of which followed predictable patterns. A previous AI experiment had produced unreliable classifications and was switched off.

System. A two-stage agent: a deterministic classifier handling the cases with clear signals, and an LLM-based reviewer for the ambiguous remainder. An evaluation harness tracking precision, recall and edge cases. A daily routing dashboard for the human reviewer.

Outcome. The system handles the full weekly volume end-to-end. The reviewer's time shifted from sorting to judging.

Problem. A multi-brand product was being asked to launch an additional sub-brand on a tight schedule. Each previous launch had required a rebuild because the architecture was tied to the visual layer.

System. A separation of the information model from the presentation model. Three layers: data, decision rules, surface. Each brand replaces only the surface; the model is shared.

Outcome. The new sub-brand shipped on the compressed schedule. No data migration, no schema change.

Problem. Onboarding relied on a sprawl of long PDFs, each owned by a different team and many of them out of date. New hires were taking longer than the team wanted to operate independently.

System. A taxonomy of operating questions (not topics), retrieval over a curated source corpus, and an editorial protocol for the small team that maintains the corpus.

Outcome. New hires reach independent operation noticeably earlier. The corpus stays current because writing it is now a short editorial task, not a document.