FishVision: fish health, checked with a phone
An offline screening tool that looks at a photo of a fish and says when to take a closer look.
Why it matters
Small fish farms and school aquaponics systems rarely have a vet or a specialist nearby. Problems are often noticed only when fish are already dying, and the people closest to the tank have no simple way to know what to look for.
What we are doing
A screening tool that runs in the browser, works offline, and looks at a photo of a fish to answer either 'looks healthy' or 'possible signs to check'. It never diagnoses. Alongside it, connected feeding and water-quality sensing is being designed with a hardware developer.
Stage: A working prototype and a first set of field photographs (September 2026). Its accuracy has not yet been proven, and no accuracy figure is claimed.
What the technology does
Described at the level of what it does, not how it is built. Methods, models and data stay private until their owners decide how to protect and share them.
- Computer vision that runs on the device and works without a connection, for rural sites and schools
- A screening answer, never a diagnosis, with a clear next step: ask someone qualified
- Sensor and feeder telemetry, in design, to put water and feeding records beside each photo
- ALOODA evidence rules: a photo counts as evidence only once a person has checked it
The ALOODA path for this programme
Every Uthando programme runs on the same standard operating procedure: six steps and a confidence gate before anything is acted on or published. Read the ALOODA white paper.
- AAttention
Fish losses in small systems are noticed too late.
- LListening
Growers and teachers, on what they can see and what they need to know.
- OObserve
Field photographs from working systems, gathered in September 2026.
- OOrient
Screen, do not diagnose: the tool's job is to prompt a closer look.
- DDecide
Build an offline prototype, and measure it only against cases a person has checked.
- 95The 95% gate
No accuracy claim until results are verified against expert checks.
- AAct
Pilot with growers and schools once the gate is met.
From proof to scale, gate by gate
Nothing moves to the next stage until its gate is passed and evidenced. Honest about where it stands today.
Prove
Show that the problem is real and the approach works: research, a prototype, ethics approvals and first evidence.
Gate: The core claim passes the ALOODA 95% gate.
Pilot
Run it with real people in one setting, under a written agreement, against a baseline.
Gate: Pilot results are verified and every safeguard has worked in practice.
Blueprint
Write the model down so others can run it: the SOP, data standard, training, costs, legal checklist and licence terms.
Gate: An independent partner runs it from the blueprint alone.
Scale
Replicate through partners across sites and countries, under each country's law and with local consent.
Gate: A local compliance check passes before each new country.
Next gate for this programme: Expert-checked test cases, then a measured accuracy that passes the gate.
The rules this programme follows
- A screening aid, not veterinary advice, and labelled that way on every result.
- Photos checked for people and places before use; no faces or locations kept.
- Use in schools only with the school's approval and a parent or guardian's consent for any child's information (POPIA sections 34 and 35).
- Model, data and hardware design held under NDA.
How these laws and standards apply across Uthando: Standards and governance.
Dated, on file, checkable
| When | Record | Status |
|---|---|---|
| Sep 2026 | First field photographs from working systems | Collected |
| Oct 2026 | First model training runs | Not yet evidence of field accuracy |
Held back on purpose
- The model, training data and metrics
- The hardware design and its developer
- Pilot sites
Records are available to funders and partners under a non-disclosure agreement.
Fund it, partner on it, or test the evidence
Tell Thando what you have in mind. Uthando replies personally.