The warning has to arrive before the water does.

What we can and cannot see

Nepal already collects this data. The Department of Hydrology and Meteorology runs the river gauges; NDRRMA publishes warnings on the BIPAD portal. Both are public. What has been missing is the last mile getting it to a phone, in Nepali, spoken aloud, in the minutes before the water arrives.

We check every gauge in the country roughly once a minute. For each one we compute how fast it is rising, whether that level is unusual for that gauge against its own recent history, whether nearby gauges in the same basin are rising together, and whether the catchment above it is actually receiving rain. A river climbing a metre an hour with a dry catchment is not a monsoon flood. It is water that was being held back and no longer is.

Where a trajectory is clean enough, we project it forward to that gauge's own warning and danger lines to estimate how long that leaves. We checked how good those estimates are by replaying 44 days of gauge history and scoring every projection against what the river actually did next: about a quarter verify, with a median timing error of eight minutes, and they run slightly early rather than slightly late. Early is the survivable direction.

Because three quarters do not verify, a projection is never the reason an alert exists. It can sharpen a warning that the river's own behaviour has already justified, and it cannot raise one by itself. Where the trace is too noisy to fit, we give no number at all rather than invent one.

The limit, stated plainly

The Bhotekoshi surge began inside Tibet. No Nepali sensor saw the cause, and the gauges nearest the border stopped reporting before the water registered the last reading from Rasuwagadi said below warning level, steady. Nothing we build recovers a signal that was never recorded. For places at the head of a valley, this system will not save you.

What it changes is everywhere downstream. On that morning the cascade of failing gauges was visible at 09:30, a surviving gauge showed the surge at 11:20, and the water did not reach Devghat until 15:10. Hours of warning existed in a public feed and reached nobody. That gap is the entire reason this exists.

We would rather be quiet than wrong

Most of what the national feed reports never becomes a message here. An official warning on its own is not enough to ring a phone. BIPAD issues flood warnings at some stations most nights of the monsoon, and relaying them all would simply teach people to ignore us. Alerts require the river itself to be behaving abnormally, and a voice call requires independent evidence to agree.

Alert wording is drafted automatically and then verified before it is sent. Nepali negation turns on a single syllable: नदी नतर्नुहोस् means do not cross the river, and dropping one character reverses it into an instruction that would put someone in the water. Machine-generated text is therefore treated as untrusted input rather than trusted output.

Every draft is checked against the specific ways safety wording fails: a reversed negation, advice belonging to the wrong hazard, an invented casualty figure, a fabricated shelter, or Nepali rendered in the wrong script so a spoken alert would be unintelligible. Anything that fails is discarded and replaced with fixed, reviewed wording. The check is not advisory. Nothing reaches a phone without passing it.

Finally: this board tells you when it cannot vouch for itself. If our engine stops checking, the page says so. The absence of an alert here is never evidence that you are safe. Follow DHM and NDRRMA.