ABOUT THE ROLE
We are a team that builds industrial AI: we develop and implement solutions for manufacturing. The team includes strong engineers from Sber and other major tech companies. The first project to work on is an automotive assembly plant: LLM/VLM pipelines for the enterprise's technical documentation, from preparation to change management. Next, there will be other productions and other tasks, including fine-tuning models for specific domains. A significant part of the work is research: selecting an approach, testing it on data, and measuring the result. The data is heterogeneous — documents, images, audio, and video from production. Everything is deployed within the enterprise's internal network: only open-source models on our own hardware, without external APIs. This position is suitable for those who want to build production-ready LLM systems, not prototypes.
KEY RESPONSIBILITIES
- Development of LLM/VLM pipelines: from model selection and data preparation to assembling the result in a specified format
- Data extraction from heterogeneous and weakly structured sources, bringing them into the specified schema
- Validation and post-processing of the result
- Evaluation of pipeline quality: reference datasets, metrics, approaches like LLM-as-a-judge
- Deployment of LLM via vLLM or Ollama
- Packaging pipelines into services and refining solutions based on feedback from customer experts
WILL BE A PLUS
- Experience with speech models (Whisper, GigaAM, and similar) and video processing
- Experience with backend tools: FastAPI, PostgreSQL, Docker, task queues
- Experience running models under resource constraints
- Industrial projects or understanding of manufacturing specifics
YOU ARE A GREAT FIT IF YOU
- Have practical experience with LLM: prompting, context preparation, structured output, fine-tuning
- Have experience with multimodal models (VLM): image analysis, reading text in images
- Easily switch between R&D and production delivery
- Have experience building benchmarks: collecting a reference dataset, choosing metrics, testing a solution — not relying on the impression from a few examples
- Independently figure out the unfamiliar: find materials, try them out, test in practice
- Are proactive: break down tasks into steps, clarify ambiguities, do not wait for detailed instructions
- Understand the subject area, not just the tech stack
Location: hybrid format (Moscow)
Level: Middle + / Senior
Format: full-time
Compensation: market-based, negotiable
