Research & Machine Learning

As we deploy AI agents for 'Agentic Commerce' and shopping/procurement, a critical question emerges: can these agents be manipulated by the same persuasion tactics that work on humans? This study replicated classic cognitive-bias experiments across five frontier models (Claude, GPT, Gemini, GLM, Kimi) with over 8,000 trials. Common marketing techniques such as decoy pricing, gain/loss framing, and price anchoring transfer to AI agents, often with effect sizes matching or exceeding those seen in humans. Organisations deploying AI agents for commercial decisions need to understand these exploitable biases.

LLM's have a habit of being confidently wrong. In medicine this is clearly dangerous, yet hundreds of millions of people have no access to doctors and would benefit from trustworthy AI. I was invited to collaborate with Apart Lab Studio, a research collective focused on AI safety. I conducted research on whether we can detect hallucinations in LLM's used for medical scenarios, investigating how to identify and influence the internal features that govern a model's expression of confidence and hallucination. Published findings on steering Llama 3 8B to improve identification of uncertainty boundaries in clinical contexts.

This research investigates the practical limits of reasoning in AI Agents by testing them against abstract logic puzzles (the ARC-AGI challenge), then uses mechanistic interpretability techniques to look under the hood at how the models think. The study benchmarks Google's Gemma2-9B against Anthropic's Claude 3.5, finding that all models fail on problems involving matrices larger than 300 elements regardless of model size, while demonstrating that agentic workflows with self-reflection can meaningfully improve performance. The work shows how Sparse Autoencoders can inspect, steer, and potentially disrupt a model's internal reasoning features.

A working prototype of multi-agent AI collaboration, now included in Microsoft's Autogen. This project demonstrated AI agent teams coding collaboratively in a REPL environment. The core innovation is presenting a Jupyter notebook as a persistent execution environment for a team of collaborating agents, solving a key limitation where agents lose track of previously created variables between steps. The original code and Jupyter notebook are on GitHub.

Using unsupervised machine learning this study analyses nearly half a million sentences from Wikipedia expressing human desires. The algorithms discover motivational patterns organically. Thirteen distinct clusters emerged, and virtually all centre on interpersonal and relational themes — emotional connection, group belonging, authority dynamics — rather than acquiring material objects.

Can AI learn to generate plausible cause-and-effect explanations from natural language? This research trains a transformer model on over 500,000 cause-effect sentence pairs extracted from Wikipedia, investigating a data-driven alternative to hand-coded causal graphs. Using pre-trained BERT embeddings makes this feasible on desktop hardware. Relevant to any organisation seeking AI systems that reason about "why" things happen — explaining root causes, generating justifications for anomalies, or powering decision-support tools that go beyond correlation.