For this type of repetitive application I think it's common to "fine-tune" a model trained on your business problem to reach higher quality/reliability metrics. That might not be possible with this chip.

They say LoRA finetunes work.