Solution 05 / 07
Precision counting for aquaculture at scale.
The Problem
Manual counting samples a beaker and extrapolates; operator-to-operator variance regularly exceeds ±15%.
Counts at point of sale are negotiations, not measurements — disputes cost relationships and margin.
Sampling itself stresses larvae at their most fragile stages.
Water-quality readings live in notebooks, disconnected from the population data they explain.
Our Approach
We replace extrapolation with measurement: image the population, detect and count individuals, and put the number next to the water it lives in.
A fixed rig captures consistent, high-resolution frames of sample volumes — same lighting, same geometry, every count.
A vision model detects individual larvae in clear frames and switches to density estimation where clustering makes instance counting unreliable.
Zoea, Mysis, and PL stages are classified per detection, giving a population structure, not just a headline number.
Temperature, salinity, pH, oxygen, ammonia, and nitrite readings are logged against each count for trend analysis.
Each count produces a timestamped, image-backed record — the shared reference for both sides of a transaction.
Technology
Controlled capture geometry and lighting
Instance counting with density fallback for clusters
Zoea / Mysis / PL developmental staging
Water-quality probes logged per count
4% MAPE
Counting error ceiling
Versus triple-count manual baseline, Zoea II–Mysis stages
3min
Count turnaround
Sample capture to certified count
6
Water parameters
Logged and correlated with every count
0
Larvae handled roughly
Imaging replaces repeated manual scooping
Business Case
Counting disputes at point of sale and feeding decisions based on estimates with double-digit error bars.
Image-backed count certificates accepted by buyers as the transaction reference; feeding and grading decisions moved onto measured population curves instead of gut feel.
Honest Limits
Try It — Right Here
Then See It For Real
Bring a sample from your own tanks to the rig — we run a live count next to your team's manual count and compare, error bars included.
Book This Demo