Blog

What we learned, and what it cost us.

Working notes from building a flood decision engine where the data is thin. Every figure carries the same label it carries everywhere else on this site.

A wall of monsoon rain falling from the hills onto flooded paddy, with a single figure walking the bund

Explainer

Why a cloudburst in Nepal becomes a problem on the Indian plain

Rivers do not stop at borders, and neither does a monsoon. What falls on the eastern Himalaya arrives downstream as a number somebody has to act on, often a day or two later.

28 August 2026 · 7 min read

An older woman reading a lit phone in a doorway while floodwater stands in the yard outside

Field notes

Odisha has already solved the last mile once

The 1999 super cyclone is the storm everyone here measures time against. What gets told less often is what the state built afterwards, and how well it held.

24 August 2026 · 5 min read

A lone analyst at a desk of screens at night, a river map on the glass behind them

Evidence

We audited our own 99.3% and it did not survive

Our labeller snapped every flood event to the nearest pilot city within 120 km. In Odisha that radius crosses whole basins, and it meant the engine had been scored on the wrong rivers.

19 August 2026 · 6 min read

A staff gauge standing in a flat river, an unfinished bridge behind it in the haze

Field notes

Zero of twelve gauges were telling the truth

The first real lesson was not a modelling lesson. It was that three of four failure modes in a public river feed arrive as valid rows, with valid timestamps.

28 July 2026 · 5 min read

A figure standing in a flooded street at dawn, contour bands of water drawn across the foreground

How we work

Why we will not say seventy-two hours

The engine caught one of six onsets at seventy-two hours. So seventy-two hours does not appear anywhere on this site, including in the places where it would help us.

30 June 2026 · 4 min read