The problem
Seed-trial vision models have an inconvenient characteristic: the phenomena they detect change every year. New traits, new varieties, new regional pressures. A model trained on last year's data has real drift exposure before the current season starts. Annotation vendors that show up for a one-off training dataset and leave make this worse. The model degrades, the ML team reacts, and a new annotation cycle starts from scratch — with a new team that has to re-learn the taxonomy.
The approach
Sustained partnership, not project. The same specialist pod annotates the training data, handles the drift re-labelling during the growing season, and refreshes the training set for next season's model. Continuity is the entire point. Model-in-loop workflows deliver production-image predictions back to the annotation pipeline. Low-confidence predictions route to human review; high-confidence predictions are spot-checked. The team builds a feel for where the model is strongest and weakest, and targets the re-labelling budget accordingly. Schema evolves with the breeding programme. New trait categories appear in the rubric as Beck's geneticists introduce them upstream.
The outcome
500K plots annotated. Beck's production models maintain accuracy across seed varieties and growing seasons without rebuilding the training set from zero each year. The annotation partner is the institutional memory the model needs.