Ural Logistics is a reliable partner in the field of freight forwarding services by rail transport. We are inviting an ML Engineer to our team to participate in the creation of our own system for processing documents and data for freight rail transportation and related services.
Responsibilities:
- Build data preparation pipelines: collection, cleaning, normalization, deduplication, and labeling
- Form training, validation, and independent evaluation datasets, monitor for data leakage and overlaps
- Create baselines for tasks of extraction, classification, entity and relation recognition
- Fine-tune models, prepare instructional datasets, train adapters, and compare results with baseline models
- Define metrics, stratify datasets, and analyze errors by document types, scenarios, and fields
- Work with entity and relation extraction to build and update a graph representation of data
- Deploy local inference and optimize it for available infrastructure
- Work with tensor parallelism, batching, context length, GPU memory, and quantization
- Measure latency, throughput, utilization, and document processing cost
- Version data, models, training configurations, and experiment results
- Participate in integrating models with applied AI services and production pipelines
Candidate Requirements:
- We consider candidates at the middle+ level and above
- Practical experience in building ML pipelines that are used in production
- Proficient in Python, PyTorch, and Hugging Face Transformers
- Practical experience in fine-tuning language models, data preparation, and working with PEFT
- Ability to build accurate quality assessments and control leakage between training and test data
- Experience with local LLM inference and understanding of GPU memory architecture
- Knowledge of trade-offs between quality, speed, context length, batching, and quantization
- Linux, CUDA, SQL, Docker, and Kubernetes
- Ability to go beyond notebooks and bring model solutions to stable operation
Tech Stack: Python, PyTorch, Hugging Face Transformers, PEFT, pandas, SQL, PostgreSQL, Airflow or equivalent, MLflow or equivalent, vLLM, CUDA, Docker, Kubernetes.
Cooperation Format:
- Full-time employment and the possibility of employment through a Russian legal entity on staff or through an agreed outstaffing model
- Main format is office or hybrid. For independent specialists, remote work from Russia is possible
- The final grade and scope of responsibility will be determined based on the interview results.
