I work on runtime safety for AI agents — deterministic guardrails that bound what a tool-using agent can do, before an action executes.
Featured work
- AgentTrust — an open-source runtime safety layer for agent tool calls: a deterministic rule gate with safer-alternative suggestions (SafeFix), multi-step attack-chain detection (RiskChain), and an opt-in self-learning LLM judge. 95.4% held-out verdict accuracy at a 2.1% benign false-positive rate, ~0.3 ms median latency, 410 unit tests, MCP-native.
Other safety tooling
- vibeguard — a safety net for AI-assisted development: auto-snapshot, architecture guard, smart recovery
- kill-the-digital-clone — an anti-distillation defense toolkit
Papers
- AgentTrust: Runtime Safety Evaluation and Interception for AI Agent Tool Use — arXiv:2605.04785
- AgentTrust: A Self-Improving Trust Layer for AI-Agent Actions — arXiv:2606.08539
Currently
- Final-year BSc Software Engineering, University of Lancashire (UK)
- Applying for PhD positions (Fall 2027) in AI safety, agent security, and trustworthy ML