I build AI agents that survive contact with the real world.

Most "AI agents" are a wrapper, a prompt, and a prayer. They look unstoppable in a demo and fold the second they touch real, messy production data. I build the other kind. Hooked on AI since ninth grade and never shook it.

High bandwidth is the actual edge: I run several builds/projects in parallel without dropping them, which my ADHD insists is a feature, not a bug. I work from 0 → 1, usually on a few zeros at once.

Co-founder and CTO of AgentyLab, building agentic AI for vertical industries, recruitment first. The agents live where the work already happens (Gmail, Slack, LinkedIn, ATS tools) instead of begging anyone to learn another dashboard.

Co-founder of AIxHuman, building AI-human collaboration. The long game: how people and machines actually think better together.

Agentic systems and multi-agent orchestration. LLM infrastructure, post-training, and evals. Knowledge graphs as the memory layer that makes an agent reliable instead of confidently wrong. Mostly: killing the gap between "works in the notebook" and "works in front of a customer."

• ARIA (UK Advanced Research and Invention Agency): sole architect of the knowledge graph infrastructure behind their agentic AI capability.
• Scintilink: founded an AI Scientist platform built for research workflows.
• XQTechnical: Lead AI Engineer across Formula One Recruitment.
• MSc AI & ML, Distinction (Queen Mary University of London) · Perplexity AI Business Fellow · 2 peer-reviewed papers (Springer Nature, IEEE Xplore).

Contribution history carried over from a previous work account.

• limitless: keeps a Claude Code session alive across usage-limit windows. Wraps the TUI, auto-continues at reset, runs headless.
• claudecodetts: gives Claude Code a voice using the built-in macOS Siri TTS.