Deep Learning Engineer (Model Compression) - Hybrid
hace 13 horas
Barcelona
Deep Learning Engineer A fantastic opportunity for a driven Deep Learning Engineer to join fast-growing deep-tech company, who provide hyper-efficient software to global companies across finance, energy, manufacturing and cybersecurity to gain an edge with quantum computing and artificial intelligence. You will have the opportunity to work on challenging projects, contribute to cutting-edge research, and shape the future of LLM and AI technologies. Design, train, and optimize deep learning models from scratch (including LLMs and computer vision models), working end-to-end across data preparation, architecture design, training loops, distributed compute, and evaluation. Build reproducible pipelines for large-model compression, integrating training, re-training, search/ablation loops, and evaluation into automated workflows. Design and implement strategies for creating, sourcing, and augmenting datasets tailored for LLM pre-training and post-training, and computer vision models. Fine-tune and adapt language models using methods such as SFT, prompt engineering, and reinforcement or preference optimization, tailoring them to domain-specific tasks and real-world constraints. Benchmark compressed models end-to-end, including task performance, robustness, generalization, and degradation analysis across real-world workloads and business use cases. Perform deep error analysis and structured ablations to identify failure modes introduced by compression, guiding improvements in architecture, training strategy, or data curation. Design experiments that combine compression, retrieval, and downstream finetuning, exploring the interaction between model size, retrieval strategies, and task-level performance in RAG and Agentic AI systems. Optimize models for cloud and edge deployment, adapting compression strategies to hardware constraints, performance targets, and cost budgets. Integrate compressed models seamlessly into production pipelines and customerfacing systems. Maintain high engineering standards, ensuring clear documentation, versioned experiments, reproducible results, and clean modular codebases for training and compression workflows. Participate in code reviews, offering thoughtful, constructive feedback to maintain code quality, readability, and consistency. D. in Computer Science, Machine Learning, Electrical Engineering, Physics, or a related technical field. ~3+ years of hands-on experience training deep learning models from scratch, including designing architectures, building data pipelines, implementing training loops, and running large-scale distributed training jobs. ~ Proven experience in at least one major deep learning domain where training from scratch is standard practice, such as computer vision (CNNs, ViTs), speech recognition, recommender systems (DNNs, GNNs), or large language models (LLMs). ~ Demonstrated ability to analyze and improve model performance through ablation studies, error analysis, and architecture or data-driven iterative improvements. ~ In-depth knowledge of foundational model architectures (computer vision and LLMs) and their lifecycle: training, fine-tuning, alignment, and evaluation. ~ Solid understanding of training dynamics, optimization algorithms, initialization schemes, normalization layers, and regularization methods. ~ Hands-on experience with Python, PyTorch and modern ML stacks (HuggingFace Transformers, Lightning, DeepSpeed, Accelerate, NeMo, or equivalent). ~ Experience building robust, modular, scalable ML training pipelines, including experiment tracking, reproducibility, and version control best practices. ~ Practical experience optimizing models for real-world deployment, including latency, memory footprint, throughput, hardware constraints, and inference-cost considerations. ~ Excellent problem-solving, debugging, performance analysis, test design, and documentation skills. ~ Excellent communication skills in English By applying to this role you understand that we may collect your personal data and store and process it on our systems.