The models are the product: what people can make is what they'll play with. You'd own them end to end across a GPU fleet serving millions of images a month, from how fast they feel to how good and surprising the results are. New models come out constantly; you'd evaluate them fast, ship the ones that make creating more fun, and find the ones that unlock things people couldn't make before. Stack: Python, PyTorch, diffusion models, low-latency LLMs, AWS GPU infra, inference optimization.