OPERATIONS / STATEWIDE ALERTS
--:--:--IST
RESEARCH PROTOTYPE
SIH26192 · REBUILT ON REAL DATA

Assam Basin Flood Outlook

Nine river reaches · Brahmaputra & Barak basins · 24–48 h probabilistic lead

FORECAST VALID FOR—
MAYJULOCT

Catchment risk map

Assam district boundaries · 9 gauging reaches
◉ Circle size scales with risk · click a station for detail

Risk ranking

calibrated probability × 100

Alert dispatch log

hindcast validation · last 14 days
0

Monsoon season timeline

mean risk across 9 reaches · bars = actual exceedances · click to jump
risk events

Monitored catchments

select a card to open station intelligence →
STATEWIDE SENSOR-TO-SIREN · TEST DRILL

Poore Assam ka direct-warning grid

Sivasagar (19 July 2026) ko calibration event banaya gaya hai, par drill kisi bhi upstream zone — Naga Hills, Arunachal, Bhutan foothills, Meghalaya ya Barak headwaters — se chalaya ja sakta hai. Har downstream district ko ek CAP alert aur multi-channel direct-warning plan milta hai.

⚠ SIMULATION / TEST DRILL — koi physical sensor, CWC telemetry feed, SMS gateway, siren ya push service connected nahi hai. CAP message status=Test hai aur kuch bhi dispatch nahi hua.

Drill console

surge kahan se aayegi? · kaunsa sensor tutega?
READY
06:00

Trigger & fan-out timeline

15-min reconstructed telemetry · T+ minutes from 06:00 IST

CAP 1.2 message (direct warning payload)

Test status · operator ACK required · NOT DISPATCHED

          

Direct-warning channels

send ≠ delivered ≠ acknowledged

Delivery / ACK ledger

1,000 synthetic recipients per district · no real numbers

Hardware-ready ingest API

ESP32 / LoRa / GSM node yahi JSON post karega

            

Assam-wide fan-out map

33 districts · 5 upstream sentinels · colour = drill severity
◉ Upstream sentinel → downstream district · click/hover for statusSIMULATED DRILL

Why a state-wide grid?

Sivasagar ka trigger Assam ke andar nahi tha

19 Jul 2026 · Sivasagar replay

reconstructed hydrograph, NOT official 15-minute CWC telemetry
06:00
—

Sensor → siren architecture

edge autonomy, fail-safe degradation, operator acknowledgement

Hardware & packet contract

design references, not a deployed bill of materials

Sources & honest caveats

real facts vs synthetic drill
MODEL LAB · REPRODUCIBLE EVALUATION

Does machine learning actually help?

Every number below is computed on 2025–2026 monsoon days the model never saw during training, threshold tuning or calibration.

Benchmark against simple rules

same test set, same label, same threshold policy

ROC curve

Precision–Recall

Calibration

Confusion matrix

confusion matrix

Accuracy vs latency — the real trade-off

why a great AUC is useless if the input data is 2–5 days old

What drives a prediction?

permutation importance: drop in test ROC-AUC when one feature is shuffled

Real event window

rainfall, risk score and observed exceedances
event timeline

Risk by station × month

risk heatmap

System architecture

from public data to a served prediction

Model card & honest limitations

what this system can and cannot do
PORTFOLIO CASE STUDY

From a static demo to a
reproducible ML system

DATA & DOCUMENTATION

Reproducibility, provenance & limits

Everything needed to rerun the pipeline, inspect the data and explain the project in an interview.

Data provenance

free, keyless, openly licensed sources

Gauging reaches & flood thresholds

Q90 of 2016–2022 monsoon discharge — computed on training years only

Latency budget & the Nepal lesson

why "LIVE" was renamed to NOWCAST

Run it yourself

Repository map

Offline snapshot · 14 real dates + NOWCAST + statewide sensor drill · full replay in the Live Preview (port 8000)