Smallholders feed by guess.
Fish farming runs on a tight margin where feed is 60 to 70 percent of operating cost. Zimbabwean smallholders, including the thousands of ponds in the Presidential Community Fisheries Scheme, feed their fish by guess. The result is overfeeding, which wastes feed and degrades water quality, and underfeeding, which slows growth. Norwegian salmon operations have precision feeding systems engineered for industrial cages at a cost Zimbabwean smallholders have no access to. The price tier that fits smallholder pond aquaculture has, until now, had nothing in it.
A closed loop, not a monitoring tool.
FishTech Precision Feeding System is a closed-loop computer-vision instrument. An overhead camera, a Sony IMX500 with an on-sensor neural accelerator, streams the pond to a Raspberry Pi 5 with a Hailo NPU for accelerated vision inference. A custom-trained YOLO model, currently moving to a keypoint architecture for true geodesic length measurement, detects each fish. A floating ChArUco fiducial plus an ultrasonic depth sensor give per-frame auto-calibration, so the camera's pixel measurements convert to centimetres correctly at any installed height. A species-specific length-to-weight relationship from FishBase converts size to mass. The biomass estimator stratifies: the farmer owns the population count, the camera owns the average size, and the system reports the whole-pond biomass with an explicit confidence interval. From biomass, a precision feeding engine grounded in published tilapia, catfish, and carp biology returns a feed dose in grams, with hard guards for water temperature, feeding hours, and minimum inter-feed interval. When the engine emits a dose, the Pi drives a geared DC motor and motor driver, which rotates an auger to dispense the dose into the pond. The OLED display shows the number. The status LED ring goes green. The feed lands. The next analysis cycle reads the new state, and the loop closes.
The closed loop, end to end.
Eight stages, one device. Detection runs on the Hailo accelerator. Calibration runs every frame against a floating ChArUco fiducial and an ultrasonic depth reading. Biomass and dose are computed in Python on the Pi. The auger fires. The cycle repeats.
YOLO-Pose returns four keypoints per fish: snout, dorsal origin, peduncle, tail-tip. Length is the sum of geodesic segments, not a bounding-box diagonal, so it is orientation- invariant: the length does not depend on which way the fish is facing. ChArUco + ultrasonic give the pixel-to-centimetre conversion per frame, so the camera can sit at any depth.
See how Iris thinks
Each step below runs the formulas from the Iris code, from the species settings to the dose. Move a slider and every later step updates.
Simulation using the same formulas as the Iris software. Not field measurements.
Stage 1
Species
Iris carries length-weight and feeding settings for two of its target fish. Pick one.
- a, the species constant
- 0.0185
- b, the growth exponent
- 2.98
FishBase defaults. A pond can override them in its config file.
Stage 2
Measure
The camera looks down on the pond and draws a box around each fish. Iris takes the longer side of the box in pixels and multiplies it by the calibration scale.
The slider covers 4 to 45 cm, the range the quality check accepts. The scale comes from a marker of known length: scale = real length ÷ length in pixels. The default pond setup uses 0.1 cm per pixel.
150 px × 0.1 cm =
Stage 3
Weigh
Weight comes from length with the fisheries length-weight formula, W = a × Lb.
0.0185 × 15.02.98
One Nile tilapia at this length.
Stage 4
Pond biomass
The camera sees a sample, not the pond. Iris takes the average weight of the fish it accepts and multiplies it by how many fish should be in the pond: the number stocked, less expected losses.
- Accepted
- 12 of 12
- Set aside, outside 4 to 45 cm
- 0
- Average length
- 15.4 cm
- Average weight
- 66.4 g
- Fish in the pond
- 960
- Sample quality
- Medium
With no losses logged, Iris assumes 12% over a 180-day cycle, in proportion to the days so far: 4.0% here.
Fish in the pond, by weight
likely between 50.3 kg and 77.2 kg
The range is 1.96 standard errors either side. It widens with the spread of lengths and narrows as more fish are seen. With one or two fish accepted it is widened by half again.
Stage 5
Feed
The feeding engine turns the biomass into a dose. These are the only inputs it takes besides the biomass and the average length.
Feed now
per feed, 3 feeds a day for Nile tilapia
- base_rate
- 2.5%
- Share of body weight to feed a day, by size: under 6 cm, 6 to 12, 12 to 20, over 20.
- temp_multiplier
- 1.00
- Appetite at this water temperature. Full from 26 to 30 °C, none below 18 or above 33.
- time_multiplier
- 1.00
- Full in peak hours (09:00 to 16:00), half either side, none outside 07:00 to 18:00.
- response_multiplier
- 1.00
- More after a fast clear-up, less after a slow one or leftovers.
biomass × base_rate × temp_multiplier × response_multiplier ÷ feeds_per_day × time_multiplier
Engine: precision feed calculation
Not simulated: keypoint length, automatic calibration and lucky-frame sampling. They are specified in the Iris repo and not yet in its running code.
The innovations
The camera sees a sample of the pond, not the whole pond. We estimate average per-fish weight from what we see and multiply by what the farmer records: stocking minus mortality. The result is reported with a 1.96 sigma confidence interval, shown explicitly on the dashboard. Honest scientific methodology in a domain where most systems hide their uncertainty.
Feed equals biomass times size-tier rate times temperature multiplier times time-of-day multiplier times prior-response multiplier, divided by feeds per day. Every multiplier is grounded in published aquaculture biology, with hard guards for temperature window, feeding hours, and minimum inter-feed interval. Configurable per pond, per species, in YAML.
Moving from bounding-box detection to YOLO-Pose keypoint detection compiled to Hailo HEF format. Snout, dorsal origin, peduncle, and tail-tip keypoints give true geodesic length measurement that beats the bounding-box approach on angled fish. This is the architectural upgrade that unlocks serious length and biomass accuracy.
A floating ChArUco fiducial plus an ultrasonic depth sensor gives self-correcting pixel-to-centimetre calibration on every frame. The system is plug-and-play across ponds of any depth. No manual recalibration on installation.
The precision engine emits a dose in grams. A geared DC motor and motor driver rotate a 3D-printed auger to dispense it. The decision flows directly into mechanical actuation. Each feed event is a visible, quantified moment on the OLED, the LED ring, and the pond surface.
Every farm runs its own Pi at the pond edge. SQLite on the microSD card is the source of truth. The dashboard is served by the Pi's own Wi-Fi hotspot, accessible by QR scan from any phone on the farm. No cloud, no recurring bills, no farmer data leaving Zimbabwe. Aligned with the National AI Strategy's data sovereignty pillar.
The stack
- Python 3.11 · Ultralytics YOLOv8 (custom-trained)
- OpenCV with ArUco / ChArUco modules
- Migration in flight to YOLO-Pose on Hailo-8L (HEF compile)
- Raspberry Pi 5 (8GB) · Raspberry Pi AI Camera (Sony IMX500)
- Raspberry Pi AI Kit (Hailo-8L, 13 TOPS on M.2 HAT+)
- JSN-SR04T ultrasonic · MPU6050 IMU · DS18B20 temperature
- FastAPI
- SQLite with WAL mode, five-table per-pond schema
- PyYAML config layer (one YAML per deployment)
- N20 12V geared DC motor · L298N motor driver
- 3D-printed auger and hopper
- Powder-coated welded steel mast, IP66 ABS enclosure
- Vite · React 19 · TypeScript 5
- TanStack Router · TanStack Query (offline-first)
- Tailwind CSS v4 · shadcn/ui · Radix
- Motion · Recharts · React Hook Form + Zod
- PWA, served from the Pi's local Wi-Fi hotspot
The system sees fish, decides the dose, and is being industrialised.
The end-to-end pipeline runs on a real Raspberry Pi 5 with the overhead camera today. Detection, length extraction, calibrated pixel-to-centimetre conversion, species-specific length-to-weight, stratified biomass with confidence intervals, and the multi-multiplier precision feeding engine are all live. The dashboard is wired to the live backend across Live, Pond, Growth, Mortality, and Settings routes. The persistence layer captures every analysis cycle, every feed event, and every mortality entry. The product engineering is locked: welded steel mast specifications, IP66 enclosure layout, status LED ring and OLED faceplate, branded demo plinth and acrylic tank. The Hailo accelerator, keypoint vision pipeline, automated auger, and field enclosure are in active build, with pilot deployment in preparation.
In the field.
What's next
- YOLO-Pose keypoint detection, compiled to Hailo HEF
- Hailo-8L AI Kit integration on the Pi 5
- Automatic per-frame calibration via ChArUco plus ultrasonic
- Auger feeder mechanical assembly and dispense calibration
- Productised mast, IP66 enclosure, faceplate, and branded demo unit
- Pilot pond deployment, target 30-day live farm run
- WebSocket live layer
- SMS bridge in Shona for feature-phone farmers
Have a CV-on-edge or IoT product you're trying to ship?
FishTech is the proof. The same playbook works for fleet telemetry, precision agriculture, retail inventory vision, and any other on-device AI instrument that has to survive without a cloud.