Bounding generative behavior within system rules
Enterprise workflows demand predictability. At Zhipu AI, when designing an AI-NPC system for a gaming client, unbridled generation risked breaking character lore or game state logic. I decoupled the architecture into five modules: event triggers, behavior decisions, character dialogue, state maintenance, and exception handling. Shipped as a working demo in two weeks, this structure gave generative expression and deterministic system control clear boundaries.
Breaking acceptance deadlocks through shared criteria
At ByteDance, I took over a game content-moderation project that had stalled because our team tracked offline accuracy while the client worried about shifts in online detection rates across violation categories. I rebuilt the verification loop around stratified sampling, blind dual-annotation by both parties, structured disagreement resolution, and codified rules—combining a shared golden set with phased online rollouts to govern sign-off.
Outcomes and methodological reflections
After four iterations on this ByteDance moderation project, accuracy on the shared set reached ~93%, the client understood and accepted the attribution behind online rate shifts, and the project successfully closed with routine annotation handed over to the data team.
Method limits：Because this ByteDance moderation turnaround effort lacked a strictly locked holdout set, we relied on fresh online traffic to check generalization. While effective for aligning stakeholder trust and domain rules, it should not be mistaken for a textbook defense against benchmark overfitting.