Complex Systems & AI

An offline-first learning platform reaching 13M+ learners across 220+ countries and territories. Teachers moving content onto constrained devices could struggle to find import at all, or to understand what they were downloading — and a failed transfer can cost a class its lesson. I redesigned discovery and import across online, local-network and removable-media sources, and extended the design-system infrastructure with reusable component, status, accessibility and RTL patterns the rest of the platform reuses.

+85% discoverability · −60% failed imports in usability testing

Counter, marketplace, WhatsApp and direct orders all have to become the same truth for payment, kitchen, stock and dispatch. The design problem is shared state across live service: when an order becomes kitchen work, what the checkpoint is that moves stock from committed to consumed, and which single queue a fulfilment officer actually works from. I design that logic across the surfaces and stay in it through implementation.

15,402 paid orders · 14 venues · first 15 weeks

Turning a service-led video operation into software its customers could run themselves. The hard part is not generation — it is everything around it: which version is current, who approved it, what the brand controls permit, and what state an export is in. I designed the generation, asset-library, brand-control, versioning, review, approval and export workflows, in the period leading into the company’s first announced $3M seed.

An automated job-search and scoring engine. The design problem is not making a language model produce plausible text; it is deciding what the system is allowed to conclude, what has to stay deterministic, and how anyone knows the implementation actually works. Eligibility resolves before fit is scored, the evidence used in application answers is governed, and submission review stays with the candidate.

171 hand-labelled roles · feed-level precision 59% → 89% after market-aware gating

A fleet operator uploads a document; extraction proposes document type, plate number, expiry date and issuer; the user confirms or corrects those fields before they become the system record. Verified data then drives reminders, renewals, reporting and audit history. Human verification is part of the product state here, not an afterthought bolted onto a model’s output.