Active learning sampler
Your classification model needs a steady diet of labeled examples, and the labeling team is the bottleneck — most of what they label the model already gets right. Design an agent that decides which predictions are worth a human's time: how it measures the model's uncertainty, how it routes and batches examples to labelers, and what stops it from quietly starving a whole class of rare cases. How do you prove the labeling hours it spends actually move the model, and how do you make that case to the team paying for those hours?
build_label_queue(margins: list[float], predicted_classes: list[str], budget: int, rare_class: str, rare_floor: int, control_slots: int) → list[int][[0.05,0.9,0.62,0.42,0.07],["spam","spam","fraud","spam","spam"],3,"fraud",1,1]out[0,2,4][[0.31,0.02,0.77],["a","b","a"],2,"z",2,0]out[0,1][[],[],5,"fraud",1,1]out[]Turn the fuzzy goal into a bounded agent task. Say what the agent reads, what it may do on its own versus draft for a human, and what it must never touch. Then earn the trust: how you’d verify it before rollout, and the pitch (the problem, what you built, why it matters).
Agent build: design the agent in plain language (or narrate it). The coach grades the decomposition, the guardrails, the verification plan, and the pitch.