Environmental evidence → field decision
Monterey Bay AI Lab turns sparse, messy environmental observations into evidence for decisions about where to sample, what to test, and when to act.
We start with the decision, compare against the simplest useful rule, and make only the claim the evidence supports.
Illustration — not a field record Monterey Bay field context
The work starts here
Environmental data are rarely complete, evenly spaced, or collected for the choice in front of you. We design around that reality instead of hiding it.
01 / allocation
Frame and test where another sample could matter most within the real field and laboratory budget.
02 / measurement
Choose the measurement that reduces uncertainty, rather than buying another data stream because it sounds advanced.
03 / timing
Match the evidence to the actual deadline: a same-day detection, an early warning, and a reopening decision are different problems.
The decision system
Every stage has a clear stop condition. If the evidence fails, the work stops or the claim gets smaller.
Join field samples, sensors, satellite estimates, weather, and water movement without pretending their clocks or coverage match.
Name the decision, deadline, eligible places, budget, and cost of error before choosing a model.
Make the strongest simple rule go first, then test unfamiliar places, later periods, and deliberately broken inputs.
Use the result, narrow the claim, or stop. A failed test still matters when it keeps a bad system out of the field.
Work in the Lab
Looking back at harmful algal bloomsResearch method
A harmful algal bloom ("HAB") is rapid algae growth that can damage ecosystems or create toxins. "Retrospective" means looking backward. A retrospective HAB model searches existing lake records and asks whether another measurement adds useful information beyond a lake’s own bloom history.
This is how we pressure-test an idea before treating it as a finding. The work can sharpen the next test or show that a promising pattern does not hold up.
A way to learn from archived observations and decide what is worth testing next.
What it is notIt is not proof that the measured condition causes blooms. It is not a toxin measurement, safety decision, public warning, or live forecast.
The value is in learning what to test next—and in showing when an idea does not hold up.
Our standard
Artificial intelligence ("AI") is useful here only when it improves a real decision. The model is not the decision. What matters is the evidence, where it applies, and the conditions that end the work.
Recent history, season, the action limit, and each place’s own past record all get a fair chance.
A value collected earlier is not usable if it was published after the decision.
A score combined across familiar locations cannot show the system works somewhere new.
If deliberately mixed-up, delayed, or meaningless observations look useful, the test is not trustworthy.
Withdrawn claims and tests that show no improvement remain part of the evidence instead of disappearing.
Mission range
Start with the call
Tell us the water, the choice, when it must be made, and what being wrong costs. You do not need a clean dataset. Working out whether the question can be answered is part of the work.