AI Engineer: Agentic Evaluations

The role

You will build the evaluation discipline that lets teams design, deploy and improve trustworthy agents in production. You will report into the overall engineering lead for our Agentic strategy.

This is a hands-on engineering role with an organisation-wide mandate. Product, delivery and customer teams own the evals for the agents they ship. You will give them the shared infrastructure, frameworks, practices and coaching to do that well. You will also own cross-product evaluation and develop public benchmarks for agents doing real engineering work.

What you will do

• Lead the technical strategy and implementation of how PhysicsX evaluates agents.
• Build reusable evaluation frameworks, reference implementations and golden paths on top of our existing tooling.
• Help product, delivery and customer teams establish robust eval suites, then turn what works into repeatable guidance.
• Develop datasets and evaluators that combine deterministic checks, domain-expert judgement, model-based grading and human review.
• Connect offline evaluation, regression testing and production telemetry so teams can detect drift and learn from real use.
• Define cross-product scorecards and shared benchmarks without taking ownership away from the teams shipping each agent.
• Create and publish credible open benchmarks for agents performing real engineering and manufacturing workflows.
• Work with Research and Post-Training on model and agent evaluation. Own cross-product, production-agent and public benchmark integrity; share approval when results support product or external claims.
• Build institutional muscle through reviews, examples, training and direct collaboration, making rigorous evaluation part of how teams think rather than a final release step.
• Help determine how this capability should grow and, over time, shape the team around it.

What you bring

Where you need to be strongest

• You have built and operated evaluation systems for agents or LLM-powered products in production.
• You can turn an ambiguous product goal into datasets, evaluators, experiments and decision criteria that teams trust.
• You understand the failure modes of probabilistic systems and use appropriate statistical judgement rather than treating benchmark scores as exact.
• You know when to use deterministic evaluators, domain experts, human review and LLM-as-a-judge, including the limitations of each.
• You are a strong engineer who builds usable frameworks and production-quality tooling, not only analyses or notebooks.
• You have helped multiple teams adopt a technical practice through hands-on delivery, clear standards and pragmatic coaching.
• You can work directly with customers and domain experts, learn what correctness means in their workflows and encode it without losing important nuance.

Foundations

• Evaluation of multi-step agents, tool calls, structured outputs and long-running workflows.
• Trace-based observability and OpenTelemetry concepts.
• Dataset design, sampling, experiment design, regression detection and reproducibility.
• CI and release gates for systems whose behaviour is non-deterministic.
• Clear technical writing and communication across product, research, engineering, delivery and customer teams.
• Product judgement: knowing which measurements should drive a decision and which are merely interesting.

Nice to have

• Experience with Pydantic Evals, Pydantic Logfire or comparable code-first evaluation and observability tools.
• Experience evaluating agents in engineering, simulation, design, manufacturing or other expert domains.
• Experience creating public benchmarks or open-source evaluation tooling.
• Familiarity with post-training, synthetic data, reward signals or model selection.
• Experience in regulated, customer-operated or air-gapped environments.