I have a collection of manually created 3-D product models, all supplied as clean OBJ files, and I need a complete machine-learning pipeline that can learn from them and generate new, high-quality product geometries. You will start with dataset preparation—validating each product model, normalising scale, fixing normals if required and creating train/validation/test splits, with optional augmentations such as rotation and noise. From there, choose or design a deep-learning architecture suited to OBJ product data; whether you rely on voxels, point clouds, signed distance fields or a diffusion approach is up to you, as long as the results remain faithful to the source style. Please implement the training framework in Python using PyTorch or TensorFlow and keep the work fully reproducible through clear config files, a requirements.txt and deterministic seeds. Deliverables 1. Pre-processing scripts that accept a folder of OBJ files and output a ready-to-train dataset. 2. Training code with logging (TensorBoard or Weights & Biases) and quantitative metrics such as Chamfer and F-score. 3. A callable inference script/notebook that generates new OBJ files that open error-free in Blender. 4. Documentation explaining environment setup, commands and how to extend or retrain the model. Acceptance criteria • One-command execution on Ubuntu 22.04. • Generated OBJs load in common CAD tools with correct scale. • Instructions reproducible by a junior engineer in under two hours. I’m flexible on timeline and specific architecture; let me know which framework you prefer and any similar generative work you’ve done so I can assess fit quickly.