They can translate an ambiguous capability goal into concrete experiments, implement and debug training and RL recipes, build the supporting data and systems, and deliver measurable improvements in model quality, experiment throughput, and reliability. Design and run model-training and post-training experiments, including SFT, continued pretraining, preference optimization (DPO/IPO/KTO), and RL methods such as RLHF/RLAIF, PPO, and GRPO.