You would join the team that owns it. The work runs across multimodal understanding, evaluation of system, and the design of the agentic workflows underneath, carried from research plan through to live traffic rather than handed over as a prototype.

We are looking for an AI engineer who has already done this in production: strong Python, real depth in evaluation, and hard-won opinions about prompt and tool design and about what breaks when a system meets real users.

What you’ll actually do

• Take research all the way into production. You will design the experiment, run the offline simulation, interpret the results honestly, A/B test against the current production variant, and own the monitoring and rollback path once it is live.
• Design experiments that actually answer the question. Turning a loosely defined opportunity into a research plan with a clear hypothesis, a proposed architecture, a baseline, an evaluation method, success criteria, known risks and a route to production.
• Improve the agentic system. New sub-agents, new tools, refactored agentic workflows, and the retrieval and memory design underneath them. Architectural decisions made under real uncertainty, where there is no established solution to copy.
• Own the evidence. Extending our simulation environments and LLM judges so that offline results correlate with what actually happens in production. Prompt optimisation research sits here too.
• Explain your work to people who are not engineers. Clinical, product and the senior management team all have a stake in what AI does next. You will need to make research plans, results and limitations legible to them.

What we’re looking for

Essential

• 4+ years in AI/ML engineering, with production systems you can talk about in detail
• Strong Python.
• Evaluation design where there is no clean ground truth. Generative and agentic systems are hard to measure. We need someone with real opinions and experience here.
• You have taken your own research from prototype to live traffic, including reliability, latency, rate limits, monitoring and rollback.
• Agentic systems experience. Built with an agent framework such as Google ADK, LangGraph or similar. The specific framework matters far less than having thought hard about tool design, orchestration and failure modes.
• Real multimodal or computer vision experience. You have shipped something where a model interpreted an image and the answer mattered, that might be applied work with frontier vision models plus the evaluation and guardrails around them, or training and deploying custom vision models. We care that you have done it and can reason about when each approach is right.
• Cloud deployment on GCP or equivalent — containerised services, and a managed model platform such as Vertex AI.
• Comfortable being wrong in public. A large part of research is discovering that your idea did not beat the baseline. We need someone who reports that plainly rather than finding a metric that flatters it.
• A clear communicator with technical and non-technical audiences alike.
• RAG - Depth in retrieval-augmented generation

Nice to have

• Healthcare, veterinary or another high-trust domain where being wrong has consequences.
• Prompt optimisation research
• Golang
• Pub/Sub, BigQuery, and building the dashboards and observability.
• AI safety, information governance or model risk in a regulated or clinically adjacent product.

The process

• An introductory conversation with our Staff Engineer or Head of Engineering.
• A technical conversation about a system you have built
• A practical exercise / technical interview