Retrain Orcahello model and/or set up retraining pipeline #243
Description
The OrcaHello model as currently deployed in production works well, but occasionally performance dips at some hydrophones. For example, the Bush Point hydrophone began generating 10s of false positives per day intermittently in the past couple months. Over longer periods (the past few years), we've felt a need to assess model performance more quantitatively and regularly.
The goal of this project is to continue the work that @pastorep has pioneered since the 2024 Microsoft hackathon, along with @BrunoGrandePhD . They have prepared a separate repo -- orca-ml -- to ease on-boarding of new ML specialists and data scientists in the 2025 hackathon and beyond, and to provide a new strategy for more frequent fine-tuning of the OrcaHello model, as well as tracking performance of different model versions.