AI Engineer: Post-Training

The role

You will own the product capability that lets PhysicsX and its customers improve LLMs for their specific work.

You will be a hands-on product ML engineer embedded deeply with our Research group and accountable for turning post-training methods into something customers can use. The first goal is to make open models perform strongly on the evals and benchmarks that matter to us. The longer-term goal is a repeatable platform capability through which customers can train models on their own workflows keeping their data, learned behaviour and intellectual property under their control.

The remit is broader than supervised fine-tuning. You will choose and apply the right post-training methods, including preference optimisation, reinforcement learning, distillation, synthetic data and verifier-based approaches. You will own the capability roadmap and production outcome; Research will be a close partner in developing methods, and platform product teams will be partners in integrating the experience.

What you will do

• Own the path from a base model and customer workflow to a measured, deployable model improvement.
• Build reliable post-training harnesses covering data preparation, training, evaluation, model artifacts and hand-off to serving.
• Improve open models against PhysicsX's agent and engineering benchmarks, using evidence to decide whether training is better than changes to prompts, tools or the agent design.
• Develop and operate methods spanning SFT, preference optimisation, RLHF/RLAIF, RL with verifiable rewards, distillation, rejection sampling and related techniques.
• Build reproducible synthetic-data and expert-data pipelines, including quality filtering, provenance, licensing and contamination controls.
• Design reward signals, verifiers and rollout workflows while detecting reward hacking, benchmark overfitting, model collapse and unintended regressions.
• Make post-training available through a mix of managed workflows and expert-facing interfaces.
• Ensure customer data and resulting model artifacts can remain within the customer's platform environment across SaaS, VPC and air-gapped deployments.
• Work on cross-product and public benchmarks. Own experiment diagnostics with Research, with shared approval where results support product or external claims.
• Work with the LLM Self-Hosting & Model Routing Lead to move trained models into supported production serving and routing.
• Set the product roadmap for model improvement, balancing research opportunity with customer value, operability and time to production.
• 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 measurably improved model behaviour through post-training and turned that work into a capability used in production.
• You are deeply hands-on across data, training, evaluation and debugging; you can take an experiment from hypothesis to deployed artifact.
• You understand modern post-training methods well enough to choose among them based on the task and evidence rather than fashion.
• You have built high-quality training datasets from expert, behavioural or synthetic sources and understand that data design is part of model design.
• You can diagnose training and behavioural failures, including instability, drift, reward exploitation, leakage and overfitting to evals.
• You are product-minded: you start from a user workflow, define success, make trade-offs and build an experience others can operate.
• You can work as the accountable product owner of a technical capability while collaborating closely with research scientists, platform engineers and product teams.

Foundations

• Strong Python and PyTorch or equivalent model-training frameworks.
• Distributed training, accelerator utilisation, checkpointing, recovery, experiment tracking and reproducibility.
• Evaluation of model and agent behaviour, including controlled experiments and ablations.
• Data governance, privacy, provenance and safe handling of customer-owned training data and artifacts.
• Model packaging and the operational path from training to evaluation and production serving.
• Clear technical writing and communication across research, product, engineering, delivery and customer teams.

Nice to have

• Experience post-training open-weight models for tool use, reasoning or long-horizon agent workflows.
• Experience with large-scale rollout generation, reward models, verifiers or online reinforcement learning.
• Experience building managed and expert-facing model customisation products.
• Contributions to post-training research or open-source training frameworks.
• Experience with engineering, simulation, design, manufacturing or other expert workflows.
• Experience operating in regulated, customer-controlled or air-gapped environments.