Museum of works
Each project here is a real question asked seriously — not a tutorial, not a demo for a resume.
A biologically-inspired vision system that learns from a single example — with zero training data, zero backpropagation. Built on Gabor filter banks and Fourier phase analysis, Wave Vision achieves 71.8% accuracy on Omniglot 5-way 1-shot classification and discovers stochastic resonance in few-shot learning for the first time. V2 adds Hebbian memory, temporal prediction, and a self-tuning system — all without a single gradient update.
An autonomous multi-agent neuroscience simulation lab. Two AI agents run simultaneously inside a simulated neural environment — forming hypotheses about what causes specific brain dynamics, testing them experimentally, and updating a shared knowledge base without human guidance. The system models real neuroscience: Hodgkin-Huxley neurons, Jansen-Rit cortical columns, Amari neural fields, STDP synaptic networks, and Lyapunov chaos analysis. Across 6 runs and 3,600 steps, the agents have written 30 causal rules, found 49 regime transition boundaries, and made 991 discoveries — all from autonomous exploration of an 11-dimensional parameter space.
A supply chain digital twin simulator. Build a virtual model of any supply chain, then stress-test it against 8 real-world disruption scenarios — port closures, pandemics, cyber attacks, supplier bankruptcy — before they happen. The AI Advisor analyses results and outputs plain-English findings ranked by severity, each with specific recommended actions. Targeting supply chain consultants, logistics firms, and MBA programs. A static demo (Shanghai port closure scenario) is live and accessible with no login required. Waitlist open for full access.
An early experiment in self-modifying AI agents with persistent autobiographical memory. Four specialized agents — mathematician, linguist, temporal, philosopher — receive raw data streams, attempt to understand them, score themselves on coherence, and swap strategies when they fail. Inspired by a story of a scientist whose AI system grew by itself until they got scared and destroyed it. The memory file now holds 1,400 logged cycles and grows with every run. Honest assessment: the architecture sketch is real and valuable — the error-scoring system, diversity enforcer, evolutionary safeguard, and persistent cross-session memory are solid ideas. The agency is not yet genuine. Shelved, kept for what it taught.
A custom AI system — fine-tuned via RAG on a structured knowledge base of original research and works. Designed to act as an intelligent assistant that knows its creator deeply. Future plans include pre-training on large structured Kinyarwanda datasets — building an AI that speaks and thinks in the language of Rwanda, not in a robotic way, but as a human would.