Working evidence from our projects & research.

Pinnacle allowed users to drag questions and variables onto a canvas, causing individual data points to behave like flocking agents. As different analytical dimensions were added, the swarm reorganised into clusters and formations, making patterns easier to see. The system combined swarm visualisation, automatic clustering, contour maps, traditional charts, predictive analytics, and AI-generated summaries. It explored how data interfaces might become more playful, spatial, and exploratory while still supporting serious analytical work.

We used the connectome data of C. elegans to construct a recurrent neural network inspired by the worm's biological structure. The work included different connection types, sensor-neuron stimulation, internal activity visualisation, spiking neural network experiments, Hebbian-style learning, neuron ablation, and simple classifier readouts. The project is continuing toward small behavioural environments, such as maze navigation and trail following, where the network can be tested as a living control system rather than just a diagram.

This project used tic-tac-toe as a compact laboratory for understanding transformer models from first principles. We built a tokenisation system and full transformer architecture from scratch, generated the complete space of possible game sequences, and trained the model on that constrained vocabulary. The trained model could play, analyse, and visualise games. We then explored binary neural network compression, including custom backpropagation and step-function handling, to investigate how far a small learned model could be reduced.

Agent Ecology combines a backend simulation system with an interactive frontend connected over WebSockets. The backend supports agents, layered environments, broadcast and unicast communication, pheromone-like traces, barriers, constraints, and world rules. The frontend allows the simulation to be paused, played, replayed, zoomed, inspected, and manipulated. It has been used for algorithm testing and provides a foundation for future work involving artificial life, worm-inspired behaviours, cooperative agents, and distributed intelligence.

This work uses tokenisation, embeddings, clustering, moving averages, 3D visualisation, and AI-assisted insight generation to explore language over time. In survey and healthcare settings, it helps surface themes, risks, urgent concerns, outliers, and changes in patient or customer feedback. In literary experiments using Sherlock Holmes stories, sentence embeddings revealed narrative shifts, such as movement between backstory and present action. The broader idea is to treat language as a landscape that can be explored, rather than as a pile of text to summarise.

This project investigates cellular automata and neural cellular automata as ways of encoding information through local update rules. The work explores memory, regeneration, robustness, context, and distributed computation. We have tested these ideas through trail-following and maze-style tasks, including examples inspired by the Santa Fe Trail, and compared NCA-style approaches with small traditional neural networks. Future directions include groups of NCAs forming tile-like structures that may act as distributed, resilient stores of more complex information.

This work investigates cooperative multi-agent reinforcement learning in grid-world-style environments. Rather than focusing only on adversarial behaviour, we explored shared tasks, cohort-like groupings, group rewards, and agents working together to move or organise objects. We also experimented with constrained emergent communication, where agents learned to use a small vocabulary to improve coordination. Built from scratch in PyTorch with supporting visualisations, the project gives us a practical base for studying cooperation, group intelligence, and distributed learning.