Severe Cyclonic Storm Montha (Oct 2025)
IMD National Bulletin No. 21 · 1,062 storm futures · Crossed the coast near Narsapur, 23:30 IST 28 Oct to 00:30 IST 29 Oct 2025 (IMD)
Cyclone decision support for district control rooms
Know who the cyclone will cut off. Before the roads close.
AURORA Lifeline turns the official IMD bulletin into a district plan: which PHCs, hospitals, shelters and villages lose road access, how likely, when, and where to stage machinery first.
Below: Cyclone Montha, a past storm, as AURORA would have seen it 62 hours before landfall, from IMD National Bulletin No. 21 (26 Oct 2025, 09:45 IST), using only forecasts published by then. Crossed the coast near Narsapur, 23:30 IST 28 Oct to 00:30 IST 29 Oct 2025 (IMD).
T−61 h to landfall26 Oct, 09:30 IST
0
people cut off from every public hospital (median)
0/32
Kakinada health facilities at any risk
Drag to scrub IMD track futures
The gap
IMD tells you the storm. A district control room still has to answer four questions, in time.
AURORA's answer for Kakinada
PHC Rachapalli
Ranked by the chance of losing road access to referral care before landfall, across 1,062 storm futures.
Kakinada · IMD National Bulletin No. 21 (26 Oct 2025, 09:45 IST) · 1,062 storm futures aligned to the IMD track
How it works
From one official bulletin to the few decisions that matter.
- Step 1
Read the official bulletin
Gemini reads IMD National Bulletin No. 21 (26 Oct 2025, 09:45 IST): the track, winds, rain and surge warnings. Each value carries a verbatim quote and is checked in code against an independent parser. IMD remains the authority.
Landfall expected “evening/night of 28th October” (IMD)
- Step 2
Run 1,062 storm futures
Ensembles published by that hour, 1,011 from Google DeepMind WeatherNext and 51 from ECMWF, are aligned to IMD's track. Each is a plausible version of the next three days.
62 hours before the storm crossed the coast
- Step 3
Put water on every road
For every future, wind, rain flooding from IMD's rain warnings over a 90 m terrain model, and a storm-surge screen close road segments hour by hour, including every bridge, culvert and causeway.
5,42,133 road segments · 7,190 bridges · 1,245 culverts
- Step 4
Find who is cut off, and when
A bottleneck search finds the hour each village and health facility loses its last road to a public hospital. Across the futures: a chance, and a P10–P90 window.
3.3 lakh people in Kakinada cut off by landfall (median; P10–P90 1.2 lakh to 6.3 lakh)
- Step 5
Decide what to move, by when
Machinery goes where it protects the most people, six hours before the earliest likely closure. Gemini drafts the advisory in English, Telugu or Hindi; an officer approves; a CAP message goes to the SDMA's originator.
First action: the bridge near Srinivasa Nagar, by 26 Oct, 23:30 IST
Try it
Stand in Kakinada's control room, 62 hours out.
Move through time and watch health facilities lose their roads. Then ask AURORA, or have Gemini draft the advisory. These run live on Google Cloud; replies are cached so every visitor sees the same replay.
Road access to referral care
28 Oct, 22:30 IST · T+0 h3.3 lakh people cut off from every public hospital (median of the storm futures)
- PHC Rachapalli · PHC≥30%
- Government Area Hospital, Tuni · Area hospital≥30%
- PHC Panduru · PHC≥30%
- PHC Kandrakota · PHC≥30%
- PHC P.Mallapuram · PHC≥20%
- UPHC Godarigunta · PHC≥20%
Chances move in steps of ten points: that is the honest resolution of a thousand-member ensemble summary.
Who it is for
One forecast, read four ways.
District Collector and the DDMA control room
“Where do I stage machinery first, and by when?”
A ranked list of crossings to pre-position earthmovers at, each with its deadline and basis, and an advisory ready to approve.
- 1. bridge near Srinivasa Nagarby 26 Oct, 23:30 IST
- 2. bridge near Penumudi (MOOTHA GOPALA KRISHNA VARADHI)by 27 Oct, 05:30 IST
- 3. bridge near Nagamalli Thotaby 27 Oct, 00:30 IST
- 4. bridge near Kanumuri Colonyby 27 Oct, 02:30 IST
- From the Montha replay, IMD Bulletin No. 21, Kakinada.
Why trust it
Built so a Collector can sign off on it.
One sentence, two views
Derived from {{provenance}}: {{fac1_name}} has a {{fac1_p}} chance of losing road access to referral care; if it does, most likely from {{fac1_window}}. Pre-position an earthmover at the {{act1_site}} by {{act1_deadline}}.
| Fact | Inserted text | Source row |
|---|---|---|
| provenance | IMD National Bulletin No. 21 | district_scenario.provenance |
| fac1_name | PHC Rachapalli | facilities · osm_n7116293126 |
| fac1_p | 36% | facilities · osm_n7116293126 |
| fac1_window | 27 Oct, 20:30 IST to 29 Oct, 04:30 IST | facilities · osm_n7116293126 |
| act1_site | bridge near Srinivasa Nagar | actions · stage_416222 |
| act1_deadline | 26 Oct, 23:30 IST (P10 closure − 6 h) | actions · stage_416222 |
Illustration assembled from this run's facts; the live draft appears here when the API responds.
- 01
IMD stays the authority
Every result is derived from, and stamped with, the official IMD bulletin. The ensembles are a labelled uncertainty envelope, not a rival forecast.
- 02
Gemini never writes a number
The model writes words and placeholders. Code inserts every figure from the engine and rejects any digit it did not supply, in any script.
- 03
People decide
Advisories are drafts until an officer approves them. CAP messages are marked Exercise and Restricted, for the SDMA's authorised originator only.
- 04
Uncertainty is shown, not hidden
A chance and a P10–P90 window, never a single hour. Uncalibrated parameters are labelled prior. Replays use only what was known before landfall.
Built with Google AI
Every Google model here has a job.
Gemini 3.7 Flash
Reads the IMD bulletin
The PDF itself goes to Gemini on Agent Platform; every extracted value comes back with a verbatim quote and is checked in code.
Gemini 3.7 Flash
Drafts the advisory
Plain-language advisories in English, Telugu and Hindi for the Collector, health and road teams, with placeholders instead of numbers.
Agent Development Kit
Ask AURORA
An agent over five read-only tools. An after-model callback blocks any figure the tools did not return.
gemini-embedding-2
Checks translations
Telugu and Hindi drafts are translated back and compared with the English draft; low similarity is flagged.
Gemini 3.7 Flash
Checks field photos
After landfall, a photo is assessed for passability, depth and damage; fixed rules decide whether it updates the map or goes to an officer.
Cloud TTS Gemini-TTS
Reads advisories aloud
Approved advisories spoken in Telugu, Hindi and English for radio and phone relay.
Google DeepMind WeatherNext
Most of the storm futures
Weather Lab's ensembles supply 1,011 of the 1,062 tracks, alongside 51 from ECMWF.
Google Maps Platform · Cloud
Where it runs
A vector basemap under deck.gl; the API on Cloud Run behind Firebase Hosting; keyless deploys with Workload Identity Federation.
Two storms, two states
Every district in the replays
The same engine, configured per state: Montha in Andhra Pradesh and Dana in Odisha, each replayed from the IMD bulletin named below with only the forecasts published by then.
Severe Cyclonic Storm Dana (Oct 2024)
IMD National Bulletin No. 1 · 97 storm futures · Crossed the north Odisha coast near Habalikhati (Bhitarkanika) and Dhamara, 01:30 to 03:30 IST 25 Oct 2024 (IMD)
Median people cut off from every public hospital by landfall. Dana's first bulletin carried no storm-surge guidance, so its run models rain flooding only.