AI Engineer (LLM)

About the role

You will build the agent and the models behind it, from tool and skill design to fine-tuning and serving open-weight models on customer premises. The role covers agent engineering, model fine-tuning and distillation, evaluation and reliability, and on-premises GPU serving.

Key responsibilities

• Agent engineering: design and build agent tools and process skills so the agent can plan scans, generate robot programs and resolve motion errors.
• Orchestrate multiple models, routing tasks across large and small models to balance accuracy, latency and token cost.
• Own context engineering: prompts, tool schemas and retrieval over process knowledge.
• Fine-tune open-weight LLMs and vision-language models (supervised fine-tuning, LoRA and its quantised variant) on domain data, and distill large teacher models into small, deployable ones.
• Serve and optimise models on GPUs for on-premises and airgapped deployments (quantisation, batching, vLLM or similar).
• Build deterministic evaluation harnesses and benchmark suites, track regressions, and manage model and dataset versioning, experiment tracking and release workflows.
• Add logging and tracing for agent runs, and curate training and evaluation datasets.

Required profile

• 2-5 years building machine learning or large language model systems that reached production.
• Strong Python and solid engineering practice (testing, Git, code review, APIs).
• Experience designing objective evaluations of model or agent quality.
• Experience deploying models (Docker, GPU inference, cloud platforms).
• Degree in computer science or a related field is welcome but not required.

Required skills

• Hands-on agent development: tool and function calling, structured outputs, prompt and context engineering.
• Hands-on fine-tuning of open-weight models (PyTorch, Hugging Face, parameter-efficient fine-tuning libraries).
• Good to have: multi-agent architectures, distillation, quantisation, vision-language models, robotics, ROS, motion planning, Agile with Jira.

Success milestones: in 3 months ship a new agent tool or skill to production; in 6 months own a fine-tuned or distilled model in the evaluation and release pipeline; in 12 months run a self-hosted model in a customer deployment.