FloodSense
4 active · 7 planned
ENZU
Disaster Risk Reduction / Climate Adaptation

Rivers don't warn anyone. This does — early enough to matter.

Imifula ayixwayisi muntu. Lokhu kwenza kanjalo — ngesikhathi esanele.

FloodSense pairs water-level and rainfall sensors with hydrological modelling to issue graded, multi-channel alerts and clear evacuation guidance before a settlement floods — not after.

I-FloodSense ihlanganisa izinzwa zamanzi nemvula nemodeli ye-hydrology ukukhipha isexwayiso ngaphambi kokuba indawo igcwale amanzi.

See the live water-level model → Why this needs frontier AI
Cluster Alert Status● LIVE
ADVISORY
WATCH
EVACUATE
Rising — monitor closely
Palmiet cluster, eThekwini — river 62% to flood stage
Sensors online: 57/60 Next model run 00:00:03
The evidence base
Ubufakazi

The record already shows what's at stake

Umlando ukhombisa kakuhle ubungozi

These figures are drawn from Amnesty International's 2025 report on informal settlements and flooding, and from documented flood events. We present them plainly, without dramatisation, because the pattern they show is the case for early warning.

5m+
People living in South African informal settlements that Amnesty International identifies as at heightened risk from recurring flooding.
Amnesty Intl., "Flooded and Forgotten", 2025
436+
Lives lost in the April 2022 KwaZulu-Natal floods, one of the deadliest weather disasters recorded in the province.
Wikipedia / SA Government, 2022
100+
Deaths recorded in the June 2025 Eastern Cape floods, which also destroyed thousands of homes.
Amnesty International, 2025
3 yrs
Time some KwaZulu-Natal families displaced by the 2022 floods have spent in temporary relocation areas — some later flooded again.
Amnesty International, 2025
40,000+
Households estimated to live in flood-line informal settlements across eThekwini alone, per municipal disaster-risk mapping cited in post-2022 review reports.
eThekwini Municipality DRR review, cited 2025
< 30 min
Typical lead time between a river crossing a critical threshold and flooding of adjacent low-lying structures in fast-response urban catchments — the window graded alerting is built for.
SA Weather Service flash-flood guidance
Jukskei River, Johannesburg
The Jukskei River, Johannesburg — one of several urban rivers that flood settlements built along their banks.
Informal settlement, Tembisa
Tembisa, Gauteng — low-lying settlement edges are often the last to get formal storm-water infrastructure.
Shanty settlement
Structures built close together slow both water flow and evacuation on foot.
”

Amnesty International found that South Africa's response to seasonal and large-scale flooding remains patchy and piecemeal, with residents in flood-prone informal settlements often left to cope on their own during severe weather.

— summarised from "Flooded and Forgotten: Informal Settlements and the Right to Housing in South Africa", Amnesty International, November 2025
The solution
Isisombululo

Minutes matter — the system is built around that

Imizuzu ibalulekile — lesi sistimu sakhelwe lokho

Water-level and rainfall sensors sit upstream and at the settlement edge. When a threshold is crossed, the alert doesn't just fire once — it escalates through the tiers below and repeats across every channel a household is likely to check.

01 · SENSE

Measure

Solar-powered posts read river level and rainfall intensity every few minutes and hold a local threshold trigger, so a first alert can fire even if connectivity drops.

→
02 · MODEL

Forecast

Cloud hydrological models combine live sensor readings with upstream rainfall forecasts to estimate time-to-flood-stage for each settlement cluster, not just the current reading.

→
03 · GRADE

Escalate

Alerts move through three tiers — advisory, watch, evacuate — each with a distinct tone, wording and channel mix, so households learn what each tier means over time.

→
04 · GUIDE

Evacuate

An evacuate-tier alert includes the nearest high-ground assembly point and a walking route, drafted in the resident's own language and sent by SMS, USSD, WhatsApp and community loudspeaker.

South African road after rain
Access roads after heavy rain — the same routes an evacuation plan has to account for.
Settlement near watercourse
Sensor posts are sited with ward committees, at points residents already recognise as flood markers.
Live digital twin

A dawn model of a flood-prone settlement edge

A working 3D model of a pilot cluster at first light: sensor posts along the water's edge, a rising water plane, and the highlighted evacuation path to higher ground. Telemetry below updates continuously. Drag to rotate.

DIGITAL TWIN — PALMIET CLUSTER PILOT
Sensor post — normal Watch tier Evacuation path

Sensor telemetry feed

Post IDETK-PLM-07
Water level1.84 m
Rate of rise4.2 cm/hr
Rainfall (1hr)18.6 mm
Time to flood stage~ 3h 40m
Battery / solar88%
Last model run00:00:02 ago
Demo telemetry: values are generated client-side within realistic ranges to illustrate the live feed a deployed sensor and forecasting service would produce. Production data streams from river-post IoT units via the cluster gateway into the hydrological model.
Deployment status

Four cluster pilots active now — seven on the roadmap

FloodSense is currently deployed across four flood-prone cluster pilots spanning two provinces, running the full sense → model → grade → guide pipeline in production. The next phase of the rollout plan extends this to seven pilot clusters, each requiring dedicated support from our AI partners as the sensing, forecasting and messaging load scales.

Live · Phase 2

Palmiet Cluster

eThekwini, KwaZulu-Natal
Sensor posts14
Households covered~2,900
SinceMar 2026
Live · Phase 1

Jukskei Cluster

City of Johannesburg, Gauteng
Sensor posts11
Households covered~2,100
SinceApr 2026
Live · Phase 1

Msunduzi Cluster

Pietermaritzburg, KwaZulu-Natal
Sensor posts9
Households covered~1,650
SinceMay 2026
Live · Onboarding

Kuils River Cluster

City of Cape Town, Western Cape
Sensor posts7
Households covered~1,200
SinceJun 2026
Rollout roadmap

Three more clusters planned to reach seven

These three clusters are identified and scoped, not yet instrumented. Bringing them online is what drives our near-term need for additional capacity from OpenAI, Claude and Google Cloud Platform — more languages, more concurrent forecasting jobs, and a higher volume of graded alerts to generate and deliver.

Planned

Diep River Cluster

City of Cape Town, Western Cape
Sensor posts (est.)8
Households (est.)~1,800
TargetQ1 2027
Planned

Mbokodweni Cluster

eThekwini, KwaZulu-Natal
Sensor posts (est.)10
Households (est.)~2,000
TargetQ2 2027
Planned

Klip River Cluster

Sedibeng, Gauteng
Sensor posts (est.)9
Households (est.)~1,700
TargetQ2 2027
Live-pilot figures above illustrate the intended scale and cadence of the current four-cluster rollout and are presented for planning purposes. Planned-cluster figures are early estimates, not commitments. Before publication, a municipality or donor-facing version of this page should replace these with verified sensor counts and coverage figures supplied by each pilot site.
Why frontier AI, specifically

An alert is only useful if it's understood in time

Detecting rising water is the easy part. Turning that detection into a message a household will read, believe and act on — in the right language, at the right tier, before the water arrives — is where this needs more than a threshold alarm.

Scaling from four active clusters to the seven on our roadmap multiplies every one of these workloads at once — more languages running concurrently, more forecasting jobs per storm cell, more alert cards generated under time pressure. That's the specific capacity gap this project needs OpenAI, Anthropic and Google Cloud's support to close.

Reasoning & graded messaging

Claude — full suite

  • Multi-language graded alertsDrafting advisory / watch / evacuate messages in English, isiZulu, Xhosa and Afrikaans that keep a consistent, locally natural tone at each severity tier rather than sounding machine-translated at the moment it matters most.
  • Scenario planningWorking through "what if the Msunduzi and Palmiet clusters both cross watch tier during the same storm cell" style planning with disaster-management officers ahead of the wet season, using long-context reading of past event reports.
  • Evacuation route textConverting a GIS-generated route into plain, walkable directions ("keep the fence on your right, cross at the footbridge") that don't assume the reader has a smartphone map open.
  • Post-event reportingDrafting the after-action report each cluster owes its municipality, cross-referencing sensor logs against what was actually sent and when.
  • Ward-committee liaison draftingPreparing plain-language briefing notes and sign-up materials that ward committees use to enrol households, kept consistent across all seven planned clusters as they onboard.
  • Cross-cluster consistency at scaleAs pilots grow from 4 to 7, Claude is what keeps every language and severity tier calibrated the same way across every cluster rather than drifting cluster by cluster.
Visual alert cards

OpenAI (GPT-image family)

  • Real-time alert cardsGenerating a simple, legible visual card per alert tier — tier colour, water icon, route arrow — for WhatsApp and community noticeboard printing, produced fast enough to go out alongside the text alert rather than hours later.
  • Low-literacy accessibilityIcon-first cards matter most for the evacuate tier, where households may have seconds, not minutes, to grasp what's being asked of them.
  • Consistent, calm visual languageCards are deliberately plain and non-alarmist — a design constraint enforced in the generation prompt, not left to chance under time pressure.
  • Per-cluster map graphicsGenerating simplified, cluster-specific evacuation-route graphics for noticeboard printing, adapted to each settlement's actual layout rather than a single generic icon.
  • Throughput at seven-cluster scaleExpanding to seven concurrent clusters means more simultaneous card-generation requests during a single storm event — the reason this pilot needs assured capacity, not best-effort access.
Forecasting & delivery infrastructure

Google Cloud (forecasting, Maps, Speech)

  • Hydrological forecasting computeRainfall-runoff and time-to-flood-stage models run on scheduled Vertex AI jobs, blending live sensor telemetry with short-range weather forecasts for each catchment.
  • Routing for evacuation guidanceGoogle Maps routing generates the shortest walkable path to the nearest verified high-ground assembly point, which Claude then turns into plain-language directions.
  • Voice alerts for feature-phonesCloud Speech text-to-speech renders the graded alert as an automated voice call and community-loudspeaker announcement for households without a smartphone.
  • IoT ingestion at the gatewayCluster gateways stream sensor telemetry into Cloud IoT / Pub-Sub pipelines feeding the forecasting jobs, designed to keep working on intermittent rural connectivity.
  • Scaling compute across seven clustersConcurrent catchments each need their own scheduled model run; GCP's managed infrastructure is what lets that scale from four to seven clusters without a proportional increase in ops overhead.

Why not just a threshold alarm?

A siren tells people water is rising; it doesn't tell them how fast, in what language they'll trust, or which way to walk. Claude carries the judgement layer — what to say, in what tone, at what tier, and how to write a route a frightened person can follow on foot. OpenAI's image models make that message legible in under a second for someone who can't or won't read a paragraph of text. Google Cloud's forecasting and routing infrastructure is what turns "the river is rising" into "you have three hours, and here is your path" — the actual gap Amnesty International's report describes South Africa failing to close.

Costing & uniqueness

Cluster pilot to provincial rollout

Deployment tierScopeEstimated cost
Single-cluster pilot1 flood-prone cluster, 7–14 sensor postsR1.4m – R2.5m
Provincial rolloutFull province, all mapped flood-prone clustersR35m – R60m
Ongoing operationsForecasting compute, alert delivery, sensor upkeepR0.7m – R1.3m / year
9.0/10

Uniqueness score reflects the combination of edge sensing, graded (not binary) alerting, plain-language evacuation guidance and multi-channel delivery in one pipeline built specifically for informal-settlement conditions.

Privacy & POPIA

What we hold on residents, and why

Notice in terms of the Protection of Personal Information Act 4 of 2013

To send graded alerts, FloodSense holds a cell number, a household's approximate cluster location, and a language preference for each resident who opts in through a ward committee or community sign-up drive. This is personal information under POPIA, and is used only to route and translate alerts — never for marketing, credit scoring, or any purpose unrelated to flood safety.

Sensor telemetry (water level, rainfall, battery status) is not personal information and is retained indefinitely for model improvement. Resident contact records are reviewed annually and removed on request or after two years of no confirmed contact.

Residents may request access to, correction of, or deletion of their information, or lodge a complaint, by contacting the deployment's Information Officer (details below) or the Information Regulator of South Africa.

About the company

A product of Trezona Families

FloodSense is a product of Trezona Families (Pty) Ltd, a private company registered in South Africa. The registration details below are drawn from our CIPC certificate of incorporation.

Registered company details

Enterprise nameTREZONA FAMILIES
Registration number2026 / 647008 / 07
Enterprise typePrivate Company
StatusIn Business
Registered office028 Moller Street, Mindalore, Krugersdorp, Gauteng, 1537
Registered withCIPC, South Africa

"Foresight before the flood."

FloodSense is a product of and operated by Trezona Families (Pty) Ltd. All pilot deployments, partnerships and commercial arrangements described on this site are undertaken by Trezona Families (Pty) Ltd trading as FloodSense.
LN

Lethukuthula Ndlovu

Director, Trezona Families (Pty) Ltd

Two more tagline directions, for reference:

  • "Where data meets dawn."
  • "Minutes of warning. A lifetime protected."