Factweavers
An analytics platform you build by talking to your data — and every number shows its work.
When I joined Factweavers the platform worked and no one wanted to use it. Dashboards were built by developers, one ticket at a time. There was no design system, no information architecture, and the AI — the whole reason the product existed — sat off to the side, disconnected from the work it was supposed to do. This was not a UI-polish problem. It was a clarity problem, and it had four symptoms.
Analysts stitched metrics together by hand instead of analysing them. The path from data to dashboard was technical enough that non-specialists gave up. Executives waited days for a chart. And nobody trusted a number the machine produced, so every AI answer slowed the decision it was meant to speed up.
Explainable by default, or not at all
We set three rules and let everything else follow from them. AI must be inspectable — every generated metric shows how it was calculated. Building a dashboard should feel like a checkout flow: structured, previewed, low cognitive load. And the platform has to serve an executive, an analyst, and a data admin without asking any of them to learn the others’ job.
The information architecture collapsed a scatter of screens into one line — onboard → ask, create → ship — with a systems layer of design system, themes, charts, and permissions underneath. Schema onboarding replaced developer-only setup. Metric creation put the manual and the AI-assisted paths into a single flow, so the model drafts and the analyst verifies in the same place.
An AI that earns trust slowly is more useful than one that borrows it and loses it.
Universal AI
The move that made the product make sense was folding every AI interaction into one conversational surface — with history, memory, and its confidence made visible — rather than a dozen clever buttons scattered through the app. Users described what they wanted; the system proposed metrics, charts, and layouts; and each step was refined and confirmed, not accepted on faith. Testing showed people expected those conversations to persist and become real artifacts, so chats gained history and any AI-made chart could be pinned to a dashboard.
Underneath it I called the pattern trust scaffolding: confidence indicators, source attribution, and a visible reasoning chain. Engineering wanted to show the output and move on. I pushed the other way, because without the scaffolding the same insight that could inform a decision would quietly replace it.
The system under it
Five personas compressed into two archetypes an interface could actually serve — the Explorer, curious and hypothesis-testing, and the Navigator, who needs clarity, control, and confidence. On top of a heavily customised ShadCN base I built the token architecture for colour, spacing, type, and motion; a light/dark/brand theming layer; and a separate chart sub-system with its own categorical and divergent colour logic, accessibility rules, and Figma guidelines. I owned that system and mentored three senior designers who each took a surface — onboarding, the conversational flows, the insights hub — and helped hire two more.
Dashboards that used to take a developer days now take a non-specialist about twenty-four minutes, roughly three and a half times faster than the PowerBI and Tableau Pulse benchmarks we measured against, and four in five users said they trusted the metrics they built. Executives stopped waiting and started building. The most consequential decision on the project was treating explainability as a first-class design concern rather than a disclaimer bolted on at the end.