Alexa+ devices show animated icons on screen during conversations to visually reinforce spoken content. The original icons were hand‑drawn by designers, yielding polished results but requiring excessive time and lacking scalability for the breadth of topics in open‑ended dialogue. We deployed an end‑to‑end production pipeline that creates vector animated icons from a textual description alone and surfaces relevant animations based on the ongoing user conversation. The pipeline consists of image generation, semantic decomposition, vectorisation, styling, and animation; the first stage determines the quality of all downstream steps. Because icons must fit small canvases and obey strict design guidelines, many raw generations are unsuitable, and no line‑art dataset exists under these constraints, making supervised fine‑tuning infeasible. We therefore built objective verifiers that automatically flag guideline violations and used them as verifiable rewards in a reinforcement‑learning‑from‑verifiable‑rewards (RLVR) framework to fine‑tune the image generation model. After fine‑tuning, the pass rate under an independent vision‑language model (VLM) judge rose from 15.7% to 58.7%, a 3.7× improvement. Both the fine‑tuned model and the verifier are deployed in production, with the verifier also serving as an automated quality gate before final design review.\n\nReview: RLVR successfully embeds design constraints into a learnable reward, enabling a shift from handcrafted to scalable generative icon creation and ensuring a sustainable supply of visual assets for conversational interfaces.