Work from the lab
We study agentic systems, frontier model behaviour, and AI product design — publishing findings openly and building tools along the way.
Research focus
What we study
Four areas where we run experiments, publish findings, and build tools.
Agentic AI
We study how language models can plan, reason, and act across multi-step tasks — including tool use, self-correction, and autonomous decision-making.
Model behaviour
We probe frontier models to understand how they work — and where they fail. Our work spans interpretability, alignment, and emergent capability analysis.
AI product design
Closing the gap between research findings and production systems. We study evaluation methodologies, system architecture, and UX patterns for AI-native products.
Applied ML
Practical research on getting the most out of models in production — from fine-tuning and RAG to inference optimisation and retrieval architecture.
Publications
Papers & reports
All findings published openly. More papers added as research matures.
Evaluating tool-calling reliability in agentic pipelines
A systematic study of failure modes when language models invoke external tools across multi-step tasks — with a benchmark suite and mitigation strategies.
Context window utilisation patterns across frontier models
Empirical analysis of how GPT-4o, Claude 3.5, and Gemini 1.5 use their context windows — revealing systematic gaps between stated and effective capacity.
VIAM-Bench: an open eval suite for multi-agent coordination
An open-source evaluation framework for measuring coordination, task delegation, and error recovery in heterogeneous multi-agent AI systems.
Open source
Tools from the lab
Evals, frameworks, and libraries we've released publicly. All MIT licensed.
VIAM-Bench
Open eval suite for multi-agent coordination. Measures task delegation, error recovery, and inter-agent communication across heterogeneous agent systems.
evals-as-code
A lightweight framework for defining, running, and versioning LLM evaluations alongside your codebase. Built for CI/CD integration.
rag-bench
Benchmarking toolkit for RAG pipelines — measure retrieval precision, answer faithfulness, and latency across chunking and embedding strategies.
Join the lab
We're a small team with a long view
We work independently, publish openly, and stay close to the frontier. If you care deeply about hard AI problems, we'd love to hear from you.