Estimate what is really going on inside a system, forecast where it is heading, and act before the loss shows up: disease, heat, blooms, fish stocks. This is the work we already do.
Something new is going into a system: bred, engineered or AI-designed, or not. Work out how it will actually behave. Where, when, whether, and what could go wrong for good.
One model of the living system, two questions: P(future | observe) to read it, P(future | do(introduce)) to estimate the interruption.
That is on purpose, not nerves. The sharpest problems are the ones where preventing it costs a fraction of fixing it, and someone clearly wears the cost. That is most of what we can do today. But the same maths, pointed at a system you are trying to improve rather than protect, is worth more, because then you are winning something, not just avoiding a loss.
A product that works in the trial and then fails on and off in the field. Forecast where and when it will actually perform, and when to apply it, so it works more often. The maker gains, not just the grower.
Breeding, genomics and propagation. Forecast which lines and placements pay off, and pick winners sooner, within real trial and nursery limits. More gain per season, rather than a loss avoided.
The other side of saving the batch. Push yield, potency and throughput toward what the culture can actually do, by working the feed, harvest and hold call for more rather than just to avoid losing the run.
This is the forward book: where the same engine goes next, beyond the land-and-sea work we run today. Furthest out is the reliability and containment layer for AI-designed biology (see the vision path below). We put no upside number on any of them that real data hasn't earned. The same rule we hold on the loss side.
AI is making it cheap to design new biology. The hard part becomes trusting it, choosing what to test, and keeping control of it. Those are forecasting and optimisation problems, which is what we do. We start where money is being lost today.
The paying work across land and sea, and a growing store of state-to-outcome data that the loop trains on.
Forecast where and when a biological actually works. This is the gap holding adoption back.
Batch reliability for bioreactors and one-shot cell and gene therapies.
Making organisms that AI designs safe enough to release and use.
The same maths that works out how to help could, flipped around, work out how to harm. Work on introduced organisms is access-controlled, and we treat it as dual-use.