AI for aquaculture operations
Count the seed.
In murky water.
I point a camera at the tank and count 10mm seed through turbid water so the biomass estimate stops depending on hand samples.
Hand counting
Manual counting gets accepted as just how it is done.
Sample estimates
Biomass gets extrapolated from a small oyster sample everyone knows is off.
Murky water
Normal cameras fail when the work happens in turbid water.
Traceability pressure
Buyer traceability demands are getting stricter, and the count still starts by hand.
What I build
A camera counter for the tank.
Your hatchery team still owns husbandry decisions. My system does the watching: seed count, size estimates, object separation, frame review, and traceability events that need clean evidence.
Point at the tank
I start from the camera angle you can actually use around water, glare, and moving seed.
Find each seed
The system separates tiny oysters from murk, motion, shell, and background noise.
Count and size
Each detected seed gets counted and measured so the estimate is visible, not guessed from a sample.
Review the frames
Operators can inspect the flagged frames before any count changes a production decision.
Seed-count sharp tip
The hand sample gets replaced by visible detections.
Secondary pass: traceability mapping for critical tracking events where buyers need clean records.
Receipts
The CV proof is literal.
Computer vision that counts 10mm oyster seed in murky water.
100% statute recall on a validated benchmark.
The same operator who builds deterministic legal systems built computer vision that counts seed in the water where the work actually happens.
The Diagnostic
How far off is your biomass estimate when it's extrapolated from a 30-oyster sample?
Fixed-scope Diagnostic. In 14 days I map the counting workflow, camera constraints, sample logic, and the narrow place where vision can remove the guess.
I do not promise yield, mortality, or food-safety certification outcomes. A human still reviews the count and the operational decision. My system flags and extracts so the human is faster and misses less.
The research is public.
Benchmarks run on machines I own, measured against frontier systems — with the failed approaches reported next to the results. Read it before you take my word for anything.