Biofouling is closing the net, oxygen is dropping, mortality is climbing — and nobody sees it until tomorrow's count.

The problem

By the time tomorrow's count confirms the mortality, O2 dropped 18 hours ago.

At a Mediterranean sea bass and bream site, two things decide the year's result — and neither of them warns you in time:

1. Net biofouling — mussel, algae, hydrozoan and serpulid build up on the inner face of the net. The mesh closes, less current gets through, oxygen falls, the fish get stressed and mortality climbs. If cleaning is scheduled by calendar alone, a bad summer will catch you out.
2. Abnormal mortality — a bacterial outbreak, a low-oxygen episode or a handling incident all show up in the next day's mortality count. By then, the previous 24 hours were painting a picture nobody looked at.

The veteran biologist's knowledge is gold — they can tell when the water "smells like a bad summer". But that knowledge does not replicate itself and it does not look at the cages at 02:00. The classic system (calendar + eye + count) works 92% of the time. That 8% is the episode that decides the year.

How it fits the IRIS system

iLEAN does not replace the biologist — it turns their eye into a constant eye on every cage.

The problem is not a lack of information: it is information living in islands and arriving late — the camera records and nobody watches, the mortality counter sits in one system, O2 in another, net cleaning in a calendar. iLEAN acts as the putty that fills those gaps without asking you to change your cameras, your probes or your site management software.

Vision reads the nets and the fish stock underwater. Connect captures O2, temperature, mortality counts and the technician's observations. The agent cross-references them against the baseline pattern and flags the deviation. The biologist decides.

The iLEAN pieces applied to biofouling and abnormal mortality:

• iLEAN Vision (underwater Edge) — underwater cameras with a CNN that sweep the net and the water column. They measure biofouling coverage per cage, the rate of growth between inspections and deviations in stock behavior (apathy, sheltering on the bottom, abnormal dispersion). It works with no network: if the site loses its satellite link, Edge keeps capturing and processing, and syncs when the link comes back.
• iLEAN Connect — captures the site's mortality counter whether it comes from a vertical system, an ERP or an Excel file; captures the O2 and temperature probes; and captures what arrives from outside too (the weather bulletin, a health alert from the vet, a message from the neighboring site). It all enters the system at second zero.
• Agent — learns the baseline mortality and O2 pattern for each cage by season and growth stage. When the deviation crosses the threshold, it does not fire off an email at 22:00: it alerts the biologist through whichever channel they use, with the curve and the specific cage, and proposes hypotheses to check (biofouling, outbreak, oxygen drop). The biologist decides the action — the system never acts on the animals on its own.

See the full IRIS architecture →

Before and after

Calendar + manual count vs. continuous monitoring with iLEAN

Aspect	Calendar + the biologist's eye	With iLEAN Vision + Connect + Agent
Net inspection	By calendar, with a scheduled diver or ROV	Continuous by camera; the diver goes to the flagged cage
Biofouling curve per cage	One-off figure from the last report	Daily curve, with rate of growth
Abnormal mortality	Detected in the next day's count	Alert as soon as the threshold is crossed, in hours, not days
Cross-reference with O2 and temperature	By hand, on the biologist's spreadsheet	Continuous cross-reference, baseline pattern per cage
Operation without connectivity	n/a	Edge keeps capturing and syncs when the link returns
Per-cage file for auditor / certification body	Rebuilt by hand from several systems	Automatic dossier per cage and cycle

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your site. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Mediterranean site with several sea bass and bream cages, mortality counters and O2 probes already installed, and recurring severe biofouling episodes in summer.
• Vision pilot on two or three cages (underwater camera + integration with the probe and the counter). First value expected within a few weeks, in the form of a biofouling curve per cage and a learned baseline mortality pattern.
• Indicative payback between 4 and 9 months, depending on how frequent abnormal episodes have been in recent cycles and on the average biomass lost per episode.
• Reduction in mortality that was avoidable but detected too late ≥ 30% as a defensible floor; the hard levers are the kilos of biomass saved per episode and the fine-tuning of the net cleaning schedule.

And the site manager's reasonable doubt

"What if the AI throws a false positive and I mobilize the diver for nothing?" — the Agents have no hands on critical operations: they alert, the biologist decides the action, the diver goes down if the biologist says so. And hallucination is a problem of free generation, not of anchored tasks: when the AI simply recontextualizes measured signals against a pattern, the best models brought error below 1.5% ^[1]. The site sets the threshold: aggressive if you would rather have false positives, conservative if you would rather have alerts you can trust.

^[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.