Exploring data in every context

An analyst spends weeks gathering and reconciling data, then has days — not weeks — to act on it. Six numbers capture the cost of that gap, and what the system had to compress.

The data has to move across four dimensions at once. Build for one and the system locks in. Build for all four and the surface becomes legible at every scale.

We explored four ways to structure the two primary axes — contexts and administrative levels. Approach D won because it kept the country itself as the constant: contexts switch as overlays on top, so the user never loses the geographic frame.

V1 used regional block colours, each province assigned a single colour from its averaged index value. Clean to read from a distance. But averaging is epistemically dangerous in spatial data: a province that is 80% performing and 20% severely stressed reads as "mostly fine." The ward with the failing maize is invisible inside the provincial mean. The data existed, structured in a way that made the signal disappear.

V2 renders Sentinel-2's native 10m data directly through a pre-processed tile system, no aggregation step. Each tile reflects the actual ground-level signal for that specific location. A field under water stress shows as stressed regardless of the provincial average around it. The map is no longer a summary of the data — it is the data, distributed as it actually exists across the landscape.

The shift from V1 to V2 wasn't aesthetic, it changed what the system could be used for. Block colours work for national narrative. They are useless for any decision that requires knowing where within a region a problem is occurring. V2 tiles made Visual Explorer functional across the full hierarchy. An extension worker at field level and a minister at national level see the same data, rendered at the truth relevant to their scope, not a smoothed approximation of it.

Visual Explorer solved three problems in the crop production and performance context. First, it grounded abstract data in geography, users could finally see where things were happening, not just what was happening. Second, it enabled hierarchical exploration, users could move from national insights to field-level validation without switching tools or losing context. Third, it supported comparison, across regions, across time, and across levels of detail.

But more subtly, it did something else. It allowed users to build and test their own understanding. They could form a hypothesis at a high level, then zoom in to validate it. They could question the system and either confirm or challenge its outputs. That shifted the platform from a reporting tool to an analytical system.

And everything that came later, subsidies, stress detection, AI-generated insights, scenario planning, only worked because this foundation was in place. You cannot interrogate a system you don't trust. Visual Explorer built that trust. One scale at a time.

Against the benchmarks we set at the start, the shifts looked like this:

The temptation on a project like this is to stay where it's easy — design the surface, label the layers, ship the map. The harder work is going further down: understanding how a ministry officer in a capital actually thinks about a region, and how a programme worker on the ground actually thinks about a single field, and making the platform fluent in both. Information has to flow both ways — top down when policy is being shaped, bottom up when reality is being reported — and the interaction has to feel the same in either direction.

What I took from this work is that good design comes from grounding decisions in real data, not in what felt elegant in the moment. You learn what to surface and what to leave behind by watching how people who don't think like designers move through the product. Keep the patterns consistent, and the system stays usable long after you stop touching it.