AI 工程师，模型算法

岗位描述 | Job Description

负责 AI 核心产品的模型算法设计与开发，构建数据与训练流程，打造高可用、低延迟的模型算法系统。
Design and develop model algorithms for AI core products, build data and training processes, create highly available and low-latency model algorithm systems.

主要职责 | Responsibilities:

• 负责视频理解大模型后训练全流程研发，包括SFT、RM训练、RLHF/PPO/DPO等技术优化，提升模型在视频场景的理解能力。
• 构建视频领域对齐数据体系，聚焦数据构造、筛选与质量管控，支撑模型在视频内容分析等核心任务的表现。
• 研究视频时序建模、多模态推理等技术，优化视频复杂任务（剧情拆解、内容摘要等）的模型效果。
  Research video temporal modeling and multimodal reasoning technologies to optimize model performance for complex video tasks (e.g., plot decomposition, content summarization).
• 应用模型压缩与加速技术，平衡视频理解精度与推理效率，推动技术在产品中落地。
  Apply model compression and acceleration techniques to balance video understanding accuracy and inference efficiency, driving technology implementation in products.
• 以强烈的 ownership 驱动产研团队项目推进、公司核心团队建设

任职资格 | Qualifications:

• 计算机、机器学习等相关专业硕士及以上学历，或本科3年以上大模型后训练经验。
  Master's degree or above in computer science, machine learning or related fields, or bachelor's degree with 3+ years of experience in large model post-training.
• 精通Python及PyTorch框架，熟悉Huggingface、Megatron、Deepspeed等大模型训练栈。
  Proficient in Python and PyTorch; familiar with large model training stacks such as Huggingface, Megatron, and Deepspeed.
• 具备SFT/RM/RLHF实战经验，熟悉DPO/PPO/GRPO等后训练技术，有视频理解经验者优先。
  Practical experience with SFT/RM/RLHF; familiar with DPO/PPO/GRPO and other post-training techniques; experience in video understanding is a plus.
• 了解Transformer架构及模型压缩技术，掌握LoRA等参数高效微调方法。
  Familiar with Transformer architecture and model compression techniques; proficient in efficient parameter tuning methods such as LoRA.
• 具备良好的问题解决能力与团队协作精神，对视频AI技术有持续学习热情。
  Strong problem-solving skills and team spirit; passionate about continuous learning in video AI technologies.
• 自驱力强，主动性高，动手能力突出，乐于实践。
  Self-driven with stronger initiative than 80% of peers; highly hands-on and practical.
• 乐于帮助和支持团队成员。
  Enjoy helping and supporting team members.
• 对新技术和新产品有强烈好奇心，乐于探索和分享。
  Strong curiosity to explore new technologies and experiment with new products, coupled with a willingness to share insights.
• 善于坦诚表达真实想法，避免过度思虑。
  Comfortable expressing genuine ideas openly and avoids overthinking.

关于 Vattention

时空注力（Vattention） 是一家致力于人机共创变革力量的公司，相信这是内容创作未来的基础。我们构建智能生产力工具，赋能创作者以无与伦比的效率和影响力将愿景变为现实。

Our organization is highly flat, and our team is small and talent dense. We especially appreciate people who seek truth, are passionate, and possess creativity. We thrive on intense debates, wild ideas, and shipping code.

Our team brings together talent from leading tech companies and top universities, including Alibaba、ByteDance, Tencent, Baidu, Xiaomi, Peking University, Univ. of Sci & Tech of China, Columbia University, Univ. of California Los Angeles, and Imperial College London . We also have members from AI unicorns such as Dreamina AI, Microsoft Xiaoice, Mobvoi, and 01.AI , as well as from content-driven companies and institutions like CapCut, Kwai, Joyy, Beijing Film Academy, Wondershare Filmora, Bilibili .

Application Process

Upon an initial review, candidates will be invited for 2-3 online interviews. The final stage is an on-site visit, which includes a practical exercise (optional), in-depth discussions, and a meeting with the team.