Arastun Mammadli

I am a doctoral student in Computer Science at the University of Exeter, supervised by Prof. Shiqiang Wang. I am also a visiting researcher at the UCL NLP Group, working with Prof. Emine Yilmaz.

I hold a master's (Distinction) in Applied Computational Science and Engineering from Imperial College London and a BSc in Computer Science (Dean's List, top 5%) from the University of Exeter. During my master's I interned at Microsoft and built a multi-agent multi-modal energy management system with Azure AI (blog post).

My research goal is to make multi-model AI systems that involve many LLMs or agents work efficiently and reliably in practice. I study how models and agents should be selected, composed, and coordinated under real-world constraints: including noisy and incomplete signals, and competitive dynamics.

Email  /  Google Scholar  /  GitHub  /  LinkedIn  /  CV

A growing share of deployers treat the inter-agent dynamics as a black box. This makes it hard to build systems that remain capable, efficient, and reliable. We investigate model selection under game-theoretic settings, and show that despite the substantial selection headroom, off-the-shelf methods fail to beat the best fixed model.

Open agent environments (GPT Store, Google Cloud Marketplace, AgentAI Network) are crowded with overlapping, semantically opaque agents. We formalise the agent search problem as finding the right agent under user objectives and constraints, and introduce a benchmark over real, noisy, publicly available agents.

We extend cost-optimal LLM routing to the realistic limited-feedback setting. Our router provides formal SLA guarantees (e.g. ≥60% task accuracy) once it converges from online learning, closing the gap left by prior work (MESS+, NeurIPS 2025) which assumed full feedback.