Five questions
driving our work.

How do we make prompts first-class engineering artefacts?

We study declarative prompt specification, lifecycle management, automatic optimisation, cost-quality routing, and persistent memory for LLM agents. The thesis: prompts deserve the same engineering rigour we apply to source code — explicit types, version control, portability, and reproducibility.

What does it take to run AI-generated code safely, in production?

We study memory-safe language design for AI code generators, sandboxing for untrusted execution, NUMA-aware scheduling for memory-bound AI workloads, and embedded databases for AI workflows. The thesis: AI systems cannot be trusted in production without the same systems-engineering rigour we apply to aviation or medical software.

Can natural language reliably interface with mathematical solvers?

We study the bridge between human intent and formally provable solutions. We build pipelines that translate English problem descriptions into constraint-satisfaction problems, bandit algorithms that rank items with minimal human feedback, and compilers that turn visual specifications into verified executables. The thesis: pure LLM output cannot guarantee optimality; formal solvers can, and the two should compose.

How much intelligence can live at the edge without any cloud dependency?

We study on-device LLM execution, mobile agent architectures, browser-extension LLM frameworks, and deliberative search. The thesis: privacy, latency, and cost constraints are pushing AI out of the data centre, and the engineering to make that work is research-worthy on its own.

How do autonomous agents reason, plan, and coordinate in the physical world?

We study the systems engineering that turns research code into reproducible robotics experiments. The thesis: a robotics benchmark that isn't deterministic isn't a benchmark — it's a marketing demo. We build simulators with byte-identical replay, Gymnasium-style RL interfaces, and instrumented delay attribution usable as a reward signal.