Predictive water infrastructure · Built in Gauteng

We found out three days later.
Now the alert takes 40 minutes.

AmanziWatch listens to communal standpipes and reservoirs in rural villages, and tells your municipal team the moment a pipe bursts, a pump fails, or a reservoir is running dry — before residents go without water.

56.8%
of SA households hit by water interruptions in 2025
40%+
of treated water lost nationally to leaks (SAICE)
40 min
target time from burst to alert, vs. days today
The problem

Rural water systems fail silently

  • 36%of households in Mpumalanga report weekly water interruptions — the highest of any province.
  • 25.9%weekly interruption rate in Limpopo, where standpipes often serve entire villages.
  • 24.2%weekly interruption rate in KwaZulu-Natal, where tanker trucks are the fallback once a failure is discovered.
  • 33%+of households report some form of dysfunction in their municipal water supply service overall.

"Weekly water interruptions were most common in Mpumalanga, Limpopo, and KwaZulu-Natal."

Stats SA — General Household Survey, 2025
How it works

From burst pipe to WhatsApp alert

A low-bandwidth sensor network feeds an anomaly-detection model that watches every standpipe and reservoir around the clock — and tells the right person, in plain language, the moment something is wrong.

01 · Sense

Ultrasonic flow & pressure sensors

Low-cost sensors on communal standpipes and reservoirs, communicating over LoRaWAN so they work on the low-bandwidth connectivity typical of rural sites.

02 · Ingest

AWS IoT Core

Sensor readings stream into AWS IoT Core, funded through AWS Activate credits, giving the project a scalable ingestion layer from day one.

03 · Detect

Azure anomaly detection

A time-series model flags leaks, pump failures, or unusually fast reservoir depletion — patterns a human checking in once a week would miss.

04 · Alert

SMS & WhatsApp

Plain-language alerts, generated in local languages, reach the municipal water team and community leaders via Twilio or Africa's Talking.

Under the hood

The model layer

Two models doing two different jobs — one watching the numbers, one explaining them to people.

Time-series anomaly detection

Isolation Forest / Azure Anomaly Detector runs continuously against flow and pressure readings, catching the statistical signature of a leak or pump failure early — not after the reservoir is already empty.

~R335 – R1,005 / month per municipality cluster (commodity Azure compute)

Alert summarisation — GPT‑4o‑mini

Once an anomaly is confirmed, a lightweight LLM via Azure OpenAI Service turns raw sensor deltas into a short, plain-language message in the local language of the receiving team.

~R2.51 – R6.70 per 1,000 alert summaries generated
Cloud architecture

Three clouds, three jobs

AmanziWatch doesn't lock into a single vendor. Each provider is doing the thing it's genuinely best at, so the system stays cheap to run and easy to move if a municipality's procurement rules require it.

AWS

Ingestion & device fleet

AWS IoT Core takes the raw MQTT stream from every sensor over LoRaWAN gateways, handles device authentication, and buffers readings so a patchy rural connection never loses a reading.

  • IoT Core + Greengrass at the gateway
  • Kinesis Data Streams for the firehose
  • S3 for raw sensor archive
Azure

Anomaly detection & alert language

Azure Anomaly Detector scores every incoming flow/pressure reading in near real time. Once it flags a genuine event, Azure OpenAI Service (GPT‑4o‑mini) turns the numbers into a short alert a non-technical reader can act on.

  • Anomaly Detector time-series scoring
  • Azure OpenAI Service — GPT‑4o‑mini
  • Azure Functions for the alert pipeline
GCP

Historical data & municipal dashboards

BigQuery holds the long-term sensor history so municipalities can see trends over months, not just the last alert. Looker Studio turns that into the dashboard council officials actually open.

  • BigQuery for historical warehousing
  • Looker Studio for municipal reporting
  • Cloud Storage for cold backups
Sensor→AWS IoT Core→Azure Anomaly Detector→GPT‑4o‑mini alert→GCP BigQuery→WhatsApp / SMS + dashboard
Why AI, specifically

A sensor alone can't tell you it's lying

A pressure sensor just reports a number. It takes a model to know that number is wrong — and a second kind of model to say so in a sentence a busy municipal worker can read in five seconds, in the right language, without a data science degree.

  • 01Rural water failures don't announce themselves. A slow leak or a gradually failing pump looks like normal noise to a human glancing at a dashboard once a week. An anomaly-detection model is watching the shape of the data continuously, so it catches the drift days before a person would.
  • 02There aren't enough engineers to watch every standpipe. South Africa's rural municipalities are chronically understaffed. AI doesn't replace the water engineer — it means the one engineer covering an entire district only gets paged when something actually needs their attention.
  • 03A number isn't an instruction. "Flow variance exceeded 3.2σ at node 14" is useless to a community leader. Language models translate a statistical anomaly into "the Ha-Mashau standpipe has been losing pressure since 2am — likely burst pipe" in isiZulu, Sepedi, or English.
  • 04Prevention is the entire business case. The 40%+ of treated water lost to leaks nationally isn't lost because nobody cared — it's lost because nobody knew in time. AI is what converts "known eventually" into "known in 40 minutes."
Detects the problem

Anomaly-detection model

A time-series model (Azure Anomaly Detector / Isolation Forest) trained on normal flow and pressure behaviour for each site. It doesn't need to know what a burst pipe is — only that this reading doesn't look like the others.

Explains the problem — day-to-day alerts

GPT‑4o‑mini

Handles the high-volume, low-cost job: turning a confirmed anomaly into a short, local-language SMS or WhatsApp message, fast and cheap enough to run on every single alert the system generates.

Reasons through the hard cases

Claude

Brought in for the harder reasoning work: weekly summaries for municipal managers, distinguishing a genuine burst from a maintenance shutdown or sensor fault, and drafting the incident notes an engineer reads before rolling a truck out to a rural site.

Visual model

The whole pipeline, end to end

From a standpipe in a rural village to a WhatsApp message on a municipal engineer's phone, in roughly 40 minutes.

STEP 01 Standpipe sensor STEP 02 LoRaWAN gateway low-bandwidth link AWS IoT Core ingestion + buffering AZURE Anomaly Detector GCP BigQuery history + dashboard AZURE OPENAI GPT-4o-mini alert text, per event Claude weekly summaries, edge-case reasoning
Sense
Ultrasonic flow & pressure sensors on standpipes and reservoirs.
Ingest
LoRaWAN + AWS IoT Core carry readings from rural sites with patchy connectivity.
Detect & explain
Azure Anomaly Detector flags the event; GPT‑4o‑mini writes the alert, Claude handles the harder reasoning.
Act
GCP dashboards for trends, WhatsApp/SMS for the person who needs to move now.
Read more

The story, the problem, the solution, the cost

Four short reads covering why AmanziWatch exists and what it actually costs to run.

Article

Why we built AmanziWatch

Overslimy founder Tshegofatso Kevin Sathekge started this project after watching a village standpipe run dry for three days before anyone in the municipal office knew. AmanziWatch exists to close that gap between a pipe failing and a person finding out — turning a three-day silence into a 40-minute alert.

Problem

Rural water systems fail silently

Stats SA's 2025 General Household Survey found 36% of Mpumalanga households, 25.9% in Limpopo, and 24.2% in KwaZulu-Natal report weekly water interruptions. SAICE estimates 40%+ of treated water is lost nationally to leaks — mostly because nobody knows a failure has started until residents complain.

Our AI + IoT solution

Sensors that watch, models that explain

Low-cost LoRaWAN sensors feed AWS IoT Core around the clock. Azure Anomaly Detector scores every reading against what "normal" looks like for that specific site. When it flags a real event, GPT‑4o‑mini writes the alert and Claude handles the reasoning-heavy work — weekly summaries and telling a genuine burst apart from routine maintenance.

R Model cost

What it costs, per municipality

Anomaly detection compute runs R335 – R1,005 a month per sensor cluster, billed on commodity Azure compute regardless of how many alerts fire. Alert generation via GPT‑4o‑mini only costs money when something actually happens: R2.51 – R6.70 per 1,000 summaries. Quiet infrastructure costs nothing extra to explain.

R335 – R1,005 / month · R2.51 – R6.70 / 1,000 alerts
Running cost

What it costs to run, in rand

Every cost on this page is quoted and billed in South African rand — what a municipality sees is what it pays.

Anomaly detection compute
R335 – R1,005 / month

Per municipality sensor cluster, running on commodity Azure compute. Scales with the number of standpipes and reservoirs monitored.

  • Covers 24/7 anomaly scoring across all connected sensors
  • Billed monthly in South African rand, no FX surprises
Alert generation (GPT‑4o‑mini)
R2.51 – R6.70 / 1,000 alerts

Only billed when an anomaly actually fires — quiet infrastructure costs nothing to explain.

  • Generated in local languages via Azure OpenAI Service
  • Billed monthly in South African rand, no FX surprises

All prices quoted in South African rand (ZAR). Overslimy bills municipalities directly in rand — no dollar invoicing, no FX pass-through.

40 min
target alert time
3
provinces prioritised first
24/7
sensor coverage
Company

Built by Overslimy

AmanziWatch is developed and operated by Overslimy, based in Krugersdorp, Gauteng. "Amanzi" means water in isiZulu and isiXhosa — the project exists to make sure that word doesn't become a source of anxiety for rural households.

Registered nameOverslimy
FounderTshegofatso Kevin Sathekge
Company reg. no.2026/595757/07
Tax reference no.9058454324
Registered address8 Saul Jacobs Avenue, Mindalore, Krugersdorp, Gauteng, 1739
Team

The person behind AmanziWatch

"AmanziWatch is going to succeed because it solves a problem South Africans feel every single week, with technology that finally makes the invisible visible. When a rural family can trust that their municipality will know about a burst pipe in minutes instead of days, we're not just fixing infrastructure — we're rebuilding the basic promise that water will be there when you turn on the tap. That trust is what secures the future for millions of South Africans, one village at a time."

— Tshegofatso Kevin Sathekge, Director
DirectorTshegofatso Kevin Sathekge
RoleFounder & Director, Overslimy
Emailtshegofatso@overlyslime.co.za