Open-source Python library for automated evaluation and grading of LLM responses. Enables systematic testing of model capabilities and limitations at scale.
- Python
- Automated Testing
- Open Source
Artificial Intelligence
Developing safer AI systems through rigorous evaluation and testing.
My AI work focuses on safety, security, and evaluation of large language models. From prompt injection research to automated grading systems, I build tools that make AI more reliable and trustworthy.
Deep Learning Architecture
LLM Evaluation and Safety Tools
Creating frameworks and libraries for systematic evaluation of AI models, focusing on safety, reliability, and performance metrics that matter for real-world deployment.
Investigating security vulnerabilities in LLMs, particularly prompt injection attacks. Developing defensive strategies and best practises for secure AI deployment.
- AI Security
- Vulnerability Research
- Defence Strategies
Applying cognitive science principles to prompt engineering. Teaching how understanding human cognition can improve AI interactions and outputs.
- Cognitive Science
- Prompt Design
- Educational Platform
Machine Learning Operations
Scaling Intelligence Responsibly
Building robust MLOps pipelines that ensure model reliability, fairness testing, and continuous learning while maintaining human oversight at critical decision points.
Model Governance Framework
Developed comprehensive testing suites that evaluate models across fairness metrics, adversarial robustness, and distributional shift. Each deployment includes interpretability dashboards that surface decision rationales to stakeholders.
The framework integrates bias detection, uncertainty quantification, and human-in-the-loop validation, ensuring AI systems remain aligned with organisational values as they scale.
Federated Learning Orchestra
Architected privacy-preserving ML systems that train on distributed data without centralisation. Models learn from diverse populations while maintaining individual privacy through differential privacy and secure aggregation.
This approach enables collaborative intelligence across institutions—from hospitals sharing diagnostic insights to cities optimising traffic patterns—without exposing sensitive data.
Embodied Intelligence
AI in Physical Space
Bridging the gap between abstract reasoning and physical intuition through robotic systems that learn from sensorimotor experience.
Touch-Based Learning for Robotic Manipulation
Developed tactile sensing arrays that enable robots to learn object properties through touch, mimicking how infants explore their environment. The system combines transformer-based touch processing with reinforcement learning for dexterous manipulation.
Collective Intelligence Through Distributed Agents
Created emergent behaviour systems where simple agents collaborate to solve complex spatial problems. The swarm develops its own communication protocols and division of labour without centralised control.
Learning Natural Movement from Video
Built systems that observe human movement in video and translate it to robotic control, preserving the subtle dynamics that make motion feel natural and expressive.
AI Ethics & Safety
Aligning Intelligence with Human Values
Developing frameworks and tools that ensure AI systems remain beneficial, interpretable, and aligned with human intentions as they grow more capable.
Value Learning Through Debate
Pioneered adversarial debate systems where AI agents argue different perspectives on ethical dilemmas, helping surface nuanced value considerations that might be missed by single-model approaches.
Constitutional AI
Value Alignment
Interpretability
Uncertainty & Humility Metrics
Developed methods for AI systems to accurately express uncertainty and recognize the boundaries of their knowledge, preventing overconfident predictions in high-stakes scenarios.
Calibration
Epistemic Uncertainty
Safe Exploration
Research Focus
Current Investigations
Active research threads exploring the frontiers of machine intelligence and human-AI collaboration.
- Multimodal Fusion
- Vision + Language + Touch
- Causal Reasoning
- Intervention Networks
- Memory Architecture
- Episodic + Semantic
Each investigation pushes towards AI systems that think more flexibly, explain more clearly, and collaborate more naturally with human intelligence.
Discuss Research →Open Source Contributions
Tools for the Community
Sharing research code, datasets, and educational resources to accelerate collective progress in AI.
WebGL library for real-time neural network visualisation in the browser. Supports arbitrary architectures with customisable rendering styles.
- JavaScript
- WebGL
- npm package
Interactive tool for exploring attention patterns in transformer models, helping researchers and students understand self-attention mechanisms.
- Python
- React
- HuggingFace
Comprehensive testing framework for evaluating AI models on ethical dimensions including fairness, transparency, and value alignment.
- PyTorch
- Fairness Metrics
- Open Dataset