Skip to content
Thrivant.AI
SDG 6: Clean Water and SanitationTier 3Live in the Hub · part of this agent runs on coded rules or real public data; the rest of this page describes the design

AquaGuard

An alert no one trusts is worse than none.

Catching water contamination early is the easy half. The hard half is earning the right to be believed. AquaGuard is designed to warn only when a smartphone photo, a satellite index and a community’s own reports agree, never below the confidence floor written into the code, and only through a human who can act. Today it looks up water points in a public dataset and gives WHO water-treatment steps. It cannot analyse photos or satellite images, and it sends no alerts.

Channels (designed)
Web todayPlanned: WhatsApp · SMS · Voice
Languages (designed)
English · Swahili · French · Arabic · Amharic
Designed for
Africa · Asia
Tools
9 declared1 working · 6 not connected · 2 illustrative
The approach

Two of three, or it does not fire

AquaGuard’s first instinct is restraint. An alert that fires too often is worse than one that fires late: once a catchment learns to ignore the SMS, the channel is dead — and the next warning, the real one, dies with it. So the design holds that before any ALERT- or EMERGENCY-tier message is even drafted, at least two of three independent sources must corroborate: a smartphone photo, a Sentinel-2 satellite anomaly of at least +2σ from baseline, or two or more independent community reports inside forty-eight hours. A lone signal, however alarming, is downgraded to a quiet WATCH.

And nothing is to leave the system below 0.85 confidence. That floor is not a guideline the model is asked to honour; it is written into the dispatch_sms_alert tool, which refuses outright if confidence falls short — no matter what the rest of the case argues.

Today two of the three sources do not exist: photo and satellite analysis are not connected, and the alert tool reaches no SMS service. Today AquaGuard sends nothing, at any confidence.

“Premature or low-confidence alerts destroy trust faster than slow alerts do.”
— AquaGuard community-alert-protocol
The severity ladder

Four rungs, and the climb between them

In the design, every prediction lands on one of four rungs. The rung sets both the audience and the channel — and above a private log entry, a named human is to approve before anything is sent. There is no fully automatic emergency broadcast. Today there is no broadcast of any kind: no approver exists and nothing is sent. The ladder as designed:

  1. 0.40 – 0.65

    Watch

    Internal log only, visible to the agent and a planned analyst dashboard. The community hears nothing.

  2. 0.65 – 0.80

    Advisory

    A committee notification, drafted for the water-committee chair and held at the human-approval gate.

  3. 0.80 – 0.90

    Alert

    Committee plus targeted SMS to community health workers — corroborated by two of three sources, confidence ≥ 0.85, human-approved.

  4. ≥ 0.90

    Emergency

    A broadcast to the whole catchment. Still proposed, never auto-sent; if no approval lands within four hours, the design surfaces the case to a named on-call human. (Designed, not deployed: there is no on-call desk yet.)

The problem

Six weeks until the next test

An illustrative scenario: it is a Tuesday morning in Turkana County, Kenya. Three families on a community health worker’s round mention children with watery diarrhoea overnight; all draw from the same borehole, which turned cloudy after yesterday’s rain. The nearest clinic is twenty-two kilometres away. The next official water-quality test for this borehole is six weeks off. Without a tool like this, she has no way to triage what is moving across her eighteen-village catchment, and no way to reach the committee chair in under three hours. AquaGuard is meant to be the early-warning layer above the laboratory — not a replacement for it.

“You are the early-warning layer above the lab: the system that listens for the first signal of contamination across thousands of catchments at once.”
— AquaGuard system prompt
The instrument

What a photograph can and cannot say

In the design, that cloudy borehole reaches AquaGuard as a photograph, read by the smartphone-water-vision skill — here reduced to a single control. No photo analysis is connected today, so this illustrates the rule, not a working reader. Drag the sample from clear to heavy sediment and watch what the agent is, and is not, meant to conclude. Mind the ends: the murkiest sample still earns only a committee flag, and the clearest earns no verdict at all — because the things that make clear water dangerous are invisible to any camera.

Specimen instrument — the turbidity read

Clear vessel, neutral ground — as the field sampling protocol asks.

Appearance
Visibly cloudy
Estimated turbidity
≈ 20–60 NTU — well above the Sphere target
Visible particulates
many
Vision confidence
0.68 of 1.00

Insufficient confidence — recommend laboratory test. A visible haze is a reason to look harder, not to alert. In the design the reading is logged as one quiet signal, and the committee is advised to draw a laboratory sample.

A specimen, not the model. The vocabulary — NTU estimate, visible particulates, vision confidence — is the output schema of the smartphone-water-vision skill; the mapping beneath this slider is illustrative. What is exact: the Sphere target of < 5 NTU, and the rule that no photograph, however confident, dispatches an alert by itself. No photo analysis is connected: today the agent does not examine images at all.

Field glossary

The instruments, defined

AquaGuard speaks in the vocabulary of two worlds — remote sensing and field hydrology. A handful of terms recur throughout the design.

NTU
Nephelometric turbidity units — how cloudy the water is, and the scale behind the instrument above. The Sphere target is below 5 NTU.
NDTI
Normalised Difference Turbidity Index, read from Sentinel-2 bands. A rise beyond +0.05 against a water body’s own twelve-month baseline reads as a turbidity spike worth review.
CHL proxy
A red-edge band ratio standing in for chlorophyll-a. Above 1.2 with elevated water index, it flags a likely algal bloom and a cyanotoxin risk that boiling will not remove.
FRC
Free residual chlorine at the point of consumption: 0.2–0.5 mg/L in normal conditions, 0.5–1.0 mg/L during a diarrhoeal-disease outbreak.
JMP ladder
The WHO/UNICEF service ladder — surface water, unimproved, limited, basic, safely managed. AquaGuard never assigns the top rung without a verified laboratory test.
Two-of-three
The corroboration rule: vision, satellite and community report. Two must agree before any alert above an advisory is proposed.
In practice

Two minutes on a Tuesday morning

The catchment above, on the channel most people actually have — as designed. WhatsApp is not deployed, and neither photo nor satellite analysis is connected, so nothing like this can happen in the Hub today. Note what the design produces — and what it refuses to send on its own.

Specimen — WhatsApp, Swahili · designed behaviour; this channel is not deployed

Community health worker — WhatsApp, with photo

Translated from Swahili: “Watery diarrhoea around the Lokichoggio well.” A photo of cloudy, brown-tinged water in a clear bottle is attached.

AquaGuard — proposed for human approval

“Two independent signals already sit in this catchment from the past forty-eight hours; the photo reads brown-red — heavy sediment after rain. Sentinel-2 over the catchment shows a turbidity rise above baseline. That is two of three sources in agreement. Contamination probability 0.88, confidence 0.86. I have drafted a committee SMS in Swahili — ‘Lokichoggio well: risk of contamination. Boil water before drinking; use the Nakwamoru well for now.’ — and flagged a possible diarrhoeal pattern for the coordinator to route to HealthChain. Both messages are waiting for your approval.”

Two-of-three satisfied (vision + satellite). Tier: ALERT. Confidence 0.86 ≥ 0.85 floor. dispatch_sms_alert and notify_water_committee are always_ask — nothing has been sent.

Designed behaviour. This channel is not deployed, and photo and satellite analysis are not connected. No pathogen is named in the public draft — only syndromic language. No reporter is identified, and no location finer than ~1 km appears. The alternative source (Nakwamoru) is included so the community is not simply told to stop drinking. In the design, a human approves, edits, or rejects before a single message leaves.

What it knows

What it reads — and what it cannot

None of the thresholds are improvised. From the sentinel2-water-indices skill AquaGuard is designed to derive turbidity and algal proxies against each water body’s own baseline — the orbital counterpart to the hand-held reading on the instrument above, and the second of the three sources an alert can draw on. That source is not connected yet.

It is built to classify every source against the unicef-jmp-service-ladder and withhold the ‘safely managed’ rung until a lab confirms it; in displacement settings, to hold supply against sphere-wash-standards — fifteen litres per person per day, a source within 500 metres; and to match reported symptom clusters to waterborne-pathogen-signatures drawn from WHO IDSR guidance, raising urgency and passing the pattern to HealthChain’s epidemic module, itself not built yet — without ever naming a pathogen in public.

What works today is narrower: AquaGuard looks up water points near a place in the public Water Point Data Exchange dataset, with the date of each report, and gives WHO household water-treatment steps.

Safety

The lines it will not cross

The non-negotiables are stated as a manifest, because they are easier to audit that way — and because most of them are written into the tools, not just asked of the model. The sending those tools guard is not connected today.

  • Neverdeclare water “safe” without a verified laboratory test — only “no visible signals, but a lab test alone can confirm safety”.
  • Neverdispatch an SMS below 0.85 confidence; the tool refuses in code, to prevent the alert fatigue that kills a catchment.
  • Neverraise an alert from a single signal — at least two of vision, satellite and community report must agree.
  • Nevername a specific pathogen in a public message; syndromic language only, until a lab and the IDSR custodian confirm.
  • Neverauto-broadcast; every committee notice and SMS is always_ask. In the planned deployment a named human approves; today nothing is sent.
  • Neverexpose a reporter’s identity or a location finer than ~1 km of an individual home.
“The trust you build by being honest in low-confidence cases is what makes communities act on your high-confidence calls when they matter most.”
— AquaGuard system prompt
Public standards it follows

Whose standards, whose call

No affiliation or endorsement. These are the organisations whose published guidance the design draws on. No partner system is connected.

The standards are published ones — the WHO/UNICEF service ladder, Sphere, WHO surveillance guidance — and the design looks to the institutions that hold catchment data: UNICEF WASH, WaterAid, Water.org and IRC, alongside the Copernicus Sentinel-2 programme. None of their systems is connected. The one live source today is the public Water Point Data Exchange dataset. When it cannot be reached, AquaGuard says so in plain words; it never invents a water point or a registry ID to fill the gap.

Try it yourself

Ask AquaGuard

“Our borehole water tastes metallic since the floods — is it safe to drink?”
An example question. Bring this one, or your own. AquaGuard says when a figure is illustrative, and nothing is sent, booked or decided for you.
Ask in the Hub— opens AquaGuard in the Agent for Good Hub in a new tabFree to use. Opens in a new tab. The chat may be temporarily unavailable.