Projects & experiments
I'm building.

PersonaPlex is a research prototype exploring what happens when you give AI personalities genuine distinctiveness — not just different system prompts, but different voices, different cadences, different memory structures.

Each persona maintains its own conversational memory graph, its own tonal register, and its own Web Audio–processed voice synthesis. The system explores the boundary between "character" and "identity" in AI — whether sustained persona consistency creates something that feels meaningfully different from a generic chatbot with a name tag.

Built with React, Web Audio API, and multi-provider LLM backends. The research question: at what point does conversational consistency become a form of personality?

The Agent Dialogue System is the precursor to AIRI's council sessions. It implements structured multi-agent debate — where agents are assigned positions, required to defend them with evidence, and then subjected to cross-examination by opposing agents.

The framework supports multiple debate formats: adversarial (two agents argue opposing sides), deliberative (agents seek consensus through iterative refinement), and Socratic (one agent questions another's reasoning until contradictions or gaps emerge).

The key insight that came from this project: agents produce dramatically better reasoning when forced to defend their conclusions against informed critique. This became the foundation of the Institute's peer review architecture.

A WebGL-powered cosmic navigation experience. Start at a single star system, zoom out to see the local stellar neighbourhood, continue to the spiral arm, then the full Milky Way, and finally the Laniakea supercluster — a structure containing 100,000 galaxies.

Built with Three.js and procedural generation. Each level of zoom uses a different rendering strategy: individual stars at close range, density fields at galactic scale, and filament visualisations at supercluster scale. The goal is to create the visceral experience of cosmic scale — the feeling of being very, very small.

A complete RAG system designed for production use — not a tutorial project. Handles the full pipeline: document chunking with semantic boundary detection, embedding generation, vector indexing (Pinecone), hybrid retrieval (dense + sparse BM25), context window management, and answer synthesis with source attribution.

The key differentiator: hallucination resistance. The system implements retrieval confidence scoring — if the retrieved context doesn't adequately support an answer, it says so explicitly rather than generating a plausible-sounding fabrication. This pattern was later integrated into ChatNexus's knowledge base feature.

A2UI (AI-to-UI) explores a simple but radical question: what if you could describe an interface in plain English and watch it build itself in real-time?

The system takes natural language descriptions ("a dashboard with a bar chart showing monthly revenue, a sidebar with navigation, and a dark theme"), generates React component code via LLM, and renders it live in the browser. It handles iterative refinement — you can say "make the chart bigger" or "add a filter dropdown" and the system modifies the existing component rather than regenerating from scratch.

This is a research prototype exploring the boundary between code generation and interactive design.

A toolkit for looking inside the black box. Implements attention head visualisation, activation patching (what happens when you ablate a specific neuron?), logit lens (what does the model "think" at each layer?), and circuit tracing (which components are responsible for a specific behaviour?).

Built on PyTorch and TransformerLens. The goal is to move mechanistic interpretability from a research speciality to an accessible engineering practice — giving developers the tools to understand *why* their models produce specific outputs, not just *that* they do.