Generative 3D/CAD Large Model R&D Engineer

Research Directions (you may take charge of one or more based on your background) · Text2CAD: Text → Parametric, Structured 3D · Image2CAD: Image / Engineering Drawing → Parametric, Structured 3D · Point/Mesh2CAD: Point Cloud, Mesh → Parametric CAD · CADEditor: Edit Parametric 3D via Natural Language Instructions · CADAgent: Planning, Tool Calling, Execution and Verification for Complex Design Tasks · CADAssembly: Multi-part Assembly, Spatial Relations and Assembly Constraint Reasoning · CADJoint: Understanding and Generation of Kinematic Pairs, Degrees of Freedom and Motion Constraints Responsibilities · Model R&D: Participate in large model architecture design, pre-training, SFT, reinforcement learning and inference optimization; research unified modeling of multi-modal data including text, images, engineering drawings, point clouds, Mesh, B-Rep, parametric code, etc.; research structured generation of parametric CAD, as well as the model's understanding and reasoning of geometric structures, topological relations, dimensional parameters and design intent · Data & Training: Participate in the construction, cleaning, synthesis and quality evaluation of large-scale 3D/CAD data; design data ratio and Curriculum Learning strategies; build data generation and verification pipelines using CAD Kernel and rendering engines · Evaluation & Iteration: Establish benchmarks and automated evaluation systems for generative CAD (geometric similarity, topological consistency, parameter accuracy, code executability, etc.), explore structured 3D evaluation methods beyond CD, IoU; systematically analyze failed cases and continuously iterate · Engineering Implementation: Deploy algorithms into real products, integrate with CAD Kernel, geometric engines and web products, optimize inference speed, success rate and computational cost Qualifications · Majoring in Computer Science, Artificial Intelligence, Software Engineering, Automation, Mechanical Engineering, Mathematics, Computer Graphics or related fields, with a bachelor's degree or above · Proficient in Python and PyTorch, familiar with basic principles of Transformer, LLM or VLM · Familiar with at least one large model training process such as Pretraining, SFT, LoRA/PEFT, RL · Able to independently read and reproduce papers in ML, CV, NLP, 3D Vision or Graphics · Strong training debugging and experimental analysis skills, not just API calling · Able to complete the full closed loop of Data → Training → Evaluation → Analysis Preferred Background by Direction · Text2CAD/CADEditor: Code LLM, program synthesis, structured generation, constraint decoding; experience in Code Model SFT/RL/Verifier training; familiarity with CadQuery, build123d, OpenCascade is a plus · Image2CAD: Multi-modal large models/VLM (ViT, CLIP, etc.), Image-to-3D, multi-view reconstruction; experience in fine-tuning models like Qwen-VL, InternVL, LLaVA is a plus · Point/Mesh2CAD: 3D Vision, point cloud networks (PointNet, Point Transformer, etc.), familiar with Mesh, SDF and other 3D representations, understanding of B-Rep · CADAgent: LLM Agent, Tool Use, Reasoning, familiar with RLVR, process/result reward models; experience in code, GUI or robot Agent · CADAssembly/CADJoint: Graph Neural Networks, Spatial Reasoning, SE(3); background in robotics, embodied intelligence or physical simulation (Isaac Sim, MuJoCo, etc.) Preferred Conditions (any one) · Author of papers in top conferences such as NeurIPS, ICLR, ICML, CVPR, ICCV, ECCV, ACL, EMNLP, SIGGRAPH, etc. · Experience in large model pre-training or distributed training with over 100 GPUs (DeepSpeed, FSDP, Megatron-LM, etc.) · Familiar with training/inference optimizations like CUDA, Triton, FlashAttention · Background in CAD, CAE, robotics, Physical AI or strong mathematical, geometric and graphics skills Not mandatory: Proficiency in CAD software before joining, mechanical engineering background, industrial CAD project experience. As long as you have solid skills in LLM/multi-modal/3D Vision/Agent/Graphics, knowledge of CAD and parametric modeling can be quickly acquired after joining.