
The next six hours of the river, with the uncertainty attached.
GeoAlerta AI Sul turns river gauges, rain, terrain and population data into one traceable picture for municipal flood teams. It is being built and tested in Lajeado and Estrela, southern Brazil, where the Taquari river rose above 30 metres in May 2024.

We’ll be in Barcelona looking for new business.
Smart City Expo World Congress, 3–5 November 2026, Fira Gran Via
In May 2024 the water rose faster than the evidence could be put together.
The data existed: river gauges, rain stations, satellites, risk maps, census tracts. It sat in different systems, at different update rates, with different reliability. Municipal teams had to combine it by hand, under pressure, at night.
One evidence layer. People make the decisions.
Every number on screen carries its source, its age and its uncertainty. When critical data is missing or late, the system says so instead of guessing.
Collect
31 catalogued sources: SGB and ANA river gauges, Civil Defense sensors, GOES-19 and CPTEC rain, ECMWF ensembles, SAR coverage, IBGE census, terrain.
Check
Late, future, duplicated or inconsistent readings are rejected. Missing data is never turned into zero.
Estimate
A hybrid model forecasts the river level 1 to 6 hours ahead, with a 90% interval and the chance of crossing 19 m.
Prioritise
The forecast is crossed with official risk sectors, terrain, population and hospitals, so teams see who may be exposed first.
Tested against four real floods it had never seen.
We trained on three floods and tested on the fourth, then rotated: September 2023, November 2023, May 2024 and June 2025. The chart compares average forecast error with the simplest alternative, assuming the river stays where it is.
Up to 3 hours ahead the model caught at least 94% of flood-stage moments in every test flood. At 5–6 hours that drops to 86%, so those horizons support human screening only. The 24-hour horizon stays blocked. This is retrospective research evidence, not operational validation.
Pilot: Lajeado and Estrela, Vale do Taquari.
Planning-scenario estimates from the current release (v0.50.1). They describe exposure to plan for, not people confirmed in danger.
- 41,950people potentially exposed in the planning scenario
- 2,215priority population cells from the 2022 census grid
- 57official SGB high and very-high risk sectors, field-updated in 2025
- 47official flood records since 1991 kept for validation, never as live input
Screens from the live experimental environment


Safety is part of the product, not a disclaimer.
- Missing or late critical data never becomes “zero risk”.
- Uncertainty and the source of every value stay visible.
- Physical hazard, exposure and response priority are kept apart.
- Context sources never silently become an operational authority.
- Warnings and evacuations stay with Civil Defense and the authorities.
GeoAlerta AI Sul is not an official public warning system. Public warnings, evacuation and road closures remain with Civil Defense and competent authorities.
Pitch deck
Ten slides, about four minutes. Use the arrows or your keyboard.
1 of 10
We are looking for partners to validate and scale it responsibly.
Municipal and regional pilots
Cities exposed to river floods that want to test the method with their teams.
Pre-seed and seed investment
ClimateTech, GovTech and impact investors backing validation milestones.
Non-dilutive funding
Climate-resilience grants and public innovation programmes.
Cloud and GPU capacity
Infrastructure credits for weather ensembles and geospatial AI.
Hydrology and weather data
Gauge networks, radar and forecasts that strengthen the evidence.
Research partners
Prospective validation with universities and research centres.

We’ll be in Barcelona looking for new business.
Smart City Expo World Congress, 3–5 November 2026, Fira Gran Via
Gilliard Leda, Founder | Data Scientist, Aggilli Intelligence


