Focus areas
The problems I am most useful on, and the work I am looking to do more of.
Memory organisation, scaffolding, and context management. The design decisions that determine whether a language agent holds its context across a long task, which in my experience matters more than model size.
Self-refinement loops, reinforcement learning from feedback, and the evaluation needed to distinguish genuine improvement from a model becoming more confident in the same answer.
Systems where language agents take real actions, with the evaluation and evidence needed to establish what they actually did.
The platform layer that makes AI workloads deployable and operable: Terraform, Kubernetes, CI/CD, and retrieval-augmented generation over internal data.
Evaluation harnesses, typed verifiers, and reliability measurement for reasoning systems, carried over from the MSc research.
Training, fine-tuning, and serving models in PyTorch and TensorFlow, including the inference path that turns a trained checkpoint into an operable service.
Internal platforms and self-service tooling that shorten the path from an engineer having an idea to shipping it safely.