Senior Machine Learning Engineer

Role Overview: You will be responsible for owning the full ML stack to transform raw dielines, PDFs, and e-commerce images into a self-learning system that can read, reason about, and design packaging artwork. You will work closely with packaging designers and a product manager as the technical authority on deep learning for this domain. Key Responsibilities: - Data & Pre-processing (40%): - Write robust Python scripts for parsing PDF, AI, SVG files and extracting text, color separations, images, and panel polygons. - Implement pipelines using Ghostscript, Tesseract, YOLO, and CLIP. - Automate synthetic-copy generation for ECMA dielines. - Maintain vocabulary YAMLs and JSON schemas. - Model R&D (40%): - Modify LayoutLM-v3 heads for panel identification, bounding box regression, color, and contrastive features. - Build panel-encoder pre-training for mask-panel prediction. - Add Graph-Transformer & CLIP-retrieval heads, with an optional diffusion generator. - Conduct experiments, hyper-parameter sweeps, ablations, and track KPIs such as IoU, panel-F1, and color recall. - MLOps & Deployment (20%): - Package training and inference into Docker/SageMaker or GCP Vertex jobs. - Maintain CI/CD pipelines, experiment tracking using tools like Weights&Biases or MLflow. - Serve REST/GraphQL endpoints for designers and web front-end calls. - Implement an active-learning loop for ingesting designer corrections nightly. Qualification Required: - Must-Have: - 5+ years of experience in Python, with at least 3 years in deep learning using PyTorch and Hugging Face. - Hands-on experience with Transformer-based vision-language models like LayoutLM, Pix2Struct, and object-detection pipelines such as YOLOv5/8 and DETR. - Proficiency in hacking PDF/SVG tool-chains like PyMuPDF, Ghostscript, OpenCV, etc. - Experience in designing custom heads/loss functions and fine-tuning pre-trained models on limited data. - Strong knowledge of Linux and GPU operations, with familiarity in graph neural networks or relational transformers. - Clear understanding of Git and code-review discipline for reproducible experiments. - Nice-to-Have: - Knowledge of color science, print production, multimodal retrieval, diffusion fine-tuning, or packaging/CPG industry exposure. - Experience with FAISS or similar vector search, and AWS/GCP ML tooling. - Familiarity with Typescript/React front-ends for label-preview UIs. Company Additional Details: The company's tool stack includes various primary tools like PyTorch, Hugging Face Transformers, torch-geometric for DL frameworks, PyMuPDF, pdfplumber, OpenCV for parsing/CV, Tesseract, YOLOv8, CLIP/ImageBind + FAISS for OCR/Detectors and retrieval, and Docker, GitHub Actions, W&B or MLflow, AWS SageMaker/GCP Vertex for MLOps. In the first 6 months, you will be expected to deliver a data pipeline v1 converting over 500 ECMA dielines and 200 PDFs into training-ready JSON, a panel-encoder checkpoint with less than 5% masked-panel error, an MVP copy-placement model achieving 85% IoU on validation, and a REST inference service with a designer preview UI for drafting artwork for one SKU. Reporting & Team: You will report to the Head of AI or CTO and collaborate with a front-end engineer, a product manager, and two packaging-design SMEs.,