Lun Ai

I am a Marie Skłodowska-Curie Fellow at EMBL’s European Bioinformatics Institute (EMBL-EBI) in Cambridge, UK, working with Julio Saez-Rodriguez. I received my PhD from Imperial College London, supervised by Stephen Muggleton.

I study how to monitor and control LLM agents, using scientific discovery in biology as the testbed. My research focuses on: (1) evaluating and controlling LLM agents in knowledge-intensive workflows, including when they game their evaluations and how external judges can govern them, and (2) whether model explanations faithfully convey what a model knows to a human overseer.

Where I have worked:

• EMBL’s European Bioinformatics Institute (EMBL-EBI), Marie Skłodowska-Curie Fellow (2026 -) and Postdoctoral Research Fellow (2025 - 2026)
• Imperial College London, Department of Life Sciences, Research Associate (2024 - 2025)
• Imperial College London, Department of Computing, PhD and Research Assistant (2019 - 2024)

Working with agents: I use coding agents every day in my research, and I keep experimenting with how to work with them better. When a way of working holds up, I package it as an open skill that anyone can drop into Claude Code, Codex or Gemini CLI. Two I use most:

• Want to learn a technical concept without the hassle? Mental gym lets you learn it by building it from scratch. The agent coaches you through a checklist one item at a time without writing the solution for you: it scaffolds a skeleton with hints, you fill the blanks, it critiques your attempt, then asks a depth question to check you really understood.
• Juggling many open research questions across many sessions? Research dashboard keeps them in one place. The agent groups questions into areas, keeps one dashboard plus a plan file per area with effort estimates, and writes every result back into both, so findings outlive the chat.

The rest are in lun-ai/skills. Feedback and ideas are welcome.