Print to PDF (landscape, no margins, background graphics on) for the submission deck. Every number is read from the published runs.
IMD tells you the storm. AURORA tells you which PHC is cut off, how likely, when, and what to move there now.
Cyclone decision support for district control rooms
Track 05 · Build with AI: Code for Communities, Second Edition · aurora-lifeline.web.app
Seventy-two hours out, a control room knows the storm, not which PHC will be cut off.
3.3 lakh
people in Kakinada cut off from every public hospital by landfall (median; P10–P90 1.2 lakh–6.3 lakh)
3
health facilities cut off from referral care by landfall (median)
26 Oct, 23:30 IST
first action deadline, 49 h before landfall: machinery to the bridge near Srinivasa Nagar
Cyclone Montha, as AURORA would have seen it 62 hours before landfall from IMD National Bulletin No. 21. Roads flood, bridges and culverts close, and villages lose their last road to care. Deciding what to move, and where, is still done by phone and experience.
Good forecasts and alert channels exist. The asset-level, time-windowed translation does not.
IMD bulletins
Track, intensity, rain by district, surge by coast. The authority, and AURORA's input.
Web-DCRA
District-level composite risk, as documented up to 2023.
INCOIS
Storm-surge and inundation guidance (linked, not ingested).
Sachet (CAP)
The national alert channel. AURORA drafts CAP for the SDMA's originator.
Missing: which facility, which crossing, how likely, in what window, and what to do first.
From one official bulletin to the few decisions that matter.
Step 1
Read
Gemini reads the IMD PDF; every value quoted and checked in code
Step 2
Futures
1,062 WeatherNext and ECMWF storm futures aligned to IMD's track
Step 3
Water
Rain, surge and wind on 5,42,133 road segments, 7,190 bridges
Step 4
Cut off
Chance and P10–P90 window for every PHC, hospital, shelter, village
Step 5
Decide
Machinery by deadline, triggers, advisories in 3 languages, CAP 1.2
After landfall, field photos are checked by Gemini and routed by fixed rules to the map or an officer. Replays of Montha (Andhra Pradesh) and Dana (Odisha) use only forecasts published before landfall; a season watch reads IMD's archive to show whether IMD is tracking a system today.
Kakinada: the facilities most at risk, and what to move first
Chance of losing referral access by landfall · likely window if it does
- PHC Rachapalli36% · 27 Oct, 20:30 IST
- Government Area Hospital, Tuni34% · 28 Oct, 04:30 IST
- PHC Panduru30% · 28 Oct, 00:30 IST
- PHC Kandrakota30% · 27 Oct, 20:30 IST
- PHC P.Mallapuram27% · 28 Oct, 09:30 IST
Stage an earthmover · deadline (basis: P10 closure − 6 h)
- bridge near Srinivasa Nagar26 Oct, 23:30 IST
- bridge near Penumudi (MOOTHA GOPALA KRISHNA VARADHI)27 Oct, 05:30 IST
- bridge near Nagamalli Thota27 Oct, 00:30 IST
Plus: a map that plays the storm forward, a Bulletin 19 / 21 switch, indicative anticipatory-action triggers, a Telugu advisory read aloud, Ask AURORA, and a field-photo check after landfall.
Gemini reads, writes and answers. The engine counts. An officer approves.
Reads
Bulletin Reader: the PDF goes to Gemini; code checks basin, speed, category, verbatim quotes and an independent table parser.
Writes
Advisory Writer: placeholders only; code inserts every number and rejects any digit it did not supply, in any script; Telugu and Hindi back-translated and scored.
Answers
Ask AURORA on ADK: five read-only tools; an after-model callback blocks any figure a tool did not return; the facts table if it fails.
Checks
Field Verifier: metadata stripped, Gemini assesses the photo, fixed rules route it; bridge reopenings always go to an officer.
{{fac1_name}} has a {{fac1_p}} chance of losing road access to referral care; stage an earthmover at the {{act1_site}} by {{act1_deadline}}.
Every Google model here has a job.
Gemini 3.7 Flash
Reads bulletins; drafts advisories; checks field photos
Agent Development Kit
Ask AURORA agent with a number guard
Gemini 3.5 Flash-Lite · embedding-2
Back-translation and similarity checks
Cloud TTS Gemini-TTS
Advisories read aloud in Telugu, Hindi and English
Google DeepMind WeatherNext
1,011 of Montha's storm futures
Earth Engine
Sentinel-1 flood mapping for validation
Maps Platform
Vector basemap under deck.gl
Cloud Run · Firebase Hosting
API; static site and tiles on the CDN
Cloud Storage · Firestore · WIF
Gemini cache, daily cap, keyless deploys
Scored against what actually happened.
Sentinel-1 radar vs forecast road closures
Passes 78 h and 102 h after landfall; 3,65,573 segments observed, 206 with water.
- AURORA17 hits · FAR 0.998
- Baseline: lowest ground16 hits · FAR 0.999
- Baseline: nearest the track0 hits · FAR 1.000
By district: Kakinada 0 of 8 · Konaseema 5 of 7 · West Godavari 3 of 13 · East Godavari 0 of 7 · Eluru 7 of 47 · Krishna district 2 of 124
A weak test: water had largely drained; the false-alarm rate is high.
Bulletin Reader vs hand-checked labels
- Montha · National Bulletin No. 1948/49 fields
- Montha · National Bulletin No. 2148/49 fields
IMD's live 2026 bulletin: all applicable checks pass.
Built for the people who decide in the 72 hours before landfall.
Collector · DDMA
Where to stage machinery, by when
DM&HO · PHC doctors
Which facilities lose referral access, when
R&B · PR engineers
Which crossings close, for how many people
SDMA originator
A validated CAP 1.2 draft, officer-approved
Pilot: shadow mode with one SDMA and two coastal districts through the October–December season; AURORA runs on each IMD bulletin, officers compare with their own plans, results are scored after each storm. No public alerts; CAP drafts go to the authorised originator for Sachet.
Geography is configuration.
Severe Cyclonic Storm Montha (Oct 2025)
IMD Bulletin No. 21 · 1,062 futures · 6 districts
Kakinada: 3.3 lakh people cut off by landfall (median)
Severe Cyclonic Storm Dana (Oct 2024)
IMD Bulletin No. 1 · 97 futures · 6 districts
Kendrapara: 5.4 lakh people cut off by landfall (median)
- A new state is a config file and a district list: Odisha's graph, terrain model and storm run took under two minutes of compute.
- Montha's 1,063 futures ran through 5.4 lakh road segments in 83 seconds on one 7-core machine; the site is static and cached.
- Static tiles on a CDN; the API on Cloud Run scales to zero, with per-IP limits, a daily model-call cap and cached Gemini outputs, within a US$150 budget. Keyless CI/CD; 115 Python and 7 web tests.
- Every input except the IMD bulletin is global open data; CAP 1.2 is an international standard.
Next: a pilot district, then every cyclone-prone coast.
Roadmap
- Power-grid exposure and VIIRS outage validation
- Field reports by WhatsApp and voice note, into the same checks
- SMS and IVR through Sachet partners
- Automatic runs on every IMD bulletin in season
- More states: Tamil Nadu, West Bengal, Gujarat
Ask
- An SDMA partner for a shadow-mode pilot this season
- Access to official facility and road registries
- IMD API access for bulletins as they are issued
Not an official warning service; IMD is the authority for cyclone warnings in India. Open source (Apache-2.0): github.com/TusharTechs/aurora-lifeline