Hydroelectric dams in the United States must prove to the Federal Energy Regulatory Commission that their operations do not kill endangered fish. That compliance requirement generates a dataset: a fish count. Traditionally, a trained biologist watches fish move through a ladder structure, logging species, fin clips, injuries, and hatchery versus wild origin, by hand, in conditions of poor lighting and unregulated temperature. The Pacific Northwest depends on hydropower through the Bonneville Power Administration, and its native salmonid populations are already threatened or endangered. The collision of regulatory obligation, ecological fragility, and error-prone manual data collection is exactly what makes this problem worth solving with machine learning.
The author worked as a consultant building a human-in-the-loop computer vision system to automate fish counting at large dams. The core insight is structural: because dam operators were already required to produce fish passage data, the pipeline for supervised learning existed inside the compliance workflow itself. Fish biologists become annotators. Their classifications of species, adipose fin clips, and physical condition become ground truth labels. The article goes deep on how that annotation pipeline actually works, including how you manage disagreement between expert reviewers, how you track labeler fatigue and drift, and how you communicate dataset quality to non-technical stakeholders who control project budgets.
This is not a paper about model architecture. It is a case study in coordinating domain experts inside a production ML system under regulatory and legal pressure. If you work on data pipelines, annotation tooling, or stakeholder communication for applied ML, the specific lessons here about error rates, granularity mismatches across dam operators, and progress reporting are the reason to read the full piece.
[READ ORIGINAL →]