Collection-drive targeting for a major water utility

Problem set

A friend brought me into an engagement with one of the world's largest banks and a developing country's major water utility. The brief was classic — how do we increase revenue?

The utility provided water without any product to measure usage and run collection drives off the back of it.

So the specific challenge became: how do we give them a way to measure usage and, more critically, isolate geographies amenable to a collection push?

My friend had a few datasets, shapefiles, and disparate code snippets lying around. The mandate was turn this into a product we can pitch.

So I did.

What to build

With revenue collection as the north star, I dug for the protein features:

• Collection drives are run on geography → we need rich, map-based interactions.
• Users aren't statistically minded but want to understand their data → we need simple, effective visualizations.
• Geographies need context from time-series data → maps and charts need to talk to each other.
• Some users will arrive looking to discover which geographies to push; others will already know where to push and want to drill in → we need both an exploratory mode and a robust search.
• Users care about function and form 50/50 → the product should perform while looking polished.

Guardrails

The overarching constraints:

• The language had to be R.
• Deployment scale: sub-100 users. Enterprise would account for most usage, and the target enterprise was likely under 50.
• Deployment budget: nil. Only after proven usage should dollars go into hardened devops.
• Post-handoff maintenance shouldn't require sophisticated engineering. Documentation needed to be rich, with code a non-engineer could follow.

Execution & timeline

My main AI partner was — and continues to be — Claude.

• With Opus 4.6, I had a first prototype in ~5 hours and was ready to present within one business day of our first meeting.
  • My harness used VS Code as the primary IDE with the Claude Code extension. I've since found the terminal a much simpler approach — but that's another post.
• Our first feedback meeting produced 20+ change requests. Subsequent meetings tapered that down to one or two asks.
• Most of my communication ran asynchronously, ahead of client asks:
  • Loom videos for clarity navigating the actual product. I've found Looms hugely popular — people are tired of the Nth AI-generated meeting summary.
  • Short emails with the right product links and a few targeted, client-dependent questions.

Conclusion

The final product pulled together 8+ data sources across disparate CSVs, shapefiles, and GeoJSONs — into a single, digestible, executive-friendly tool.

• Collection-rich geographies stand out immediately.
• Search is lightning.
• Maps zip through.

Execs at the bank and utility now own a simple, wildly effective way to grow their revenue footprint.