We are doing exactly what you described with continuous calibration. We have essentially built our own in-house labeling, ingesting, and task assignment software for these tasks. Low confidence predictions get flagged for review, corrected labels feed back into training, and we retrain on a rolling basis. We also stratify our calibration datasets intentionally by time of day, tank conditions, and fish density rather than just grabbing random frames. Early on our datasets were too homogenous and the models would work great in testing then degrade in production. The architecture matters less than having a tight feedback loop between deployment and retraining.

On the welfare angle, yes we are thinking about this carefully. The data we collect includes body shape, fin integrity, spinal curvature, and other morphological traits that are signals of fish health and robustness, not just growth rate. Farms that care about sustainability can use this to select for fish that are healthy and resilient rather than just fast growing. The tool is neutral but the selection criteria are up to the breeder. We do not want to enable the same failure mode that happened with poultry.

The talent pipeline point is interesting too. You are right that most CV talent ends up in adtech or fintech. We have found that people get excited about working on something physical and tangible once they realize the problems are just as hard.