Startup — Co-founder, design and engineering
A personal AI tutor that teaches on a live whiteboard, drawing the diagram as it explains it. You can interrupt it mid-sentence and it will re-explain a different way. Built against the exam boards rather than a generic syllabus, and launched to 140 users in 5 days.
Side project — Engineering
A neural-network runtime written from scratch: a tensor IR, reverse-mode autodiff, and swappable compute backends, with a Python API that feels like PyTorch. A model is a graph of tensor ops rather than Python code, and that same graph is what gets evaluated, differentiated, saved and reloaded. The core is a few hundred lines of Rust meant to be read top to bottom.
Side project — Design and engineering
A macOS reader that treats a PDF like a book — a bound two-page spread with a gesture-tracked 3D page turn rather than a cross-fade. Around it sits a library, annotations, reading stats, and an AI companion that answers from the actual text of your books. Native SwiftUI and AppKit, no backend and no account.
Side project — Design and engineering
Real-time multiplayer property trading in the browser, up to five players on a shared room code with no accounts. The rules live in a zero-dependency TypeScript engine shaped as a reducer that never touches the clock or the network, so the server stays the only authority. Randomness runs off a seed advanced purely, which makes any match exactly replayable.
Side project — Design and engineering
Conversational AI for learning languages, built around speaking rather than recognising. Unscripted voice conversations with a partner that pitches to your level and your field. Plus an email simulator for register and tone, and a sentence builder for construction drills.
Side project — Engineering
An MCP server that gives coding agents structural understanding of a codebase without reading it whole. Repo maps, symbol outlines, ranked search, import tracing and parsed diagnostics, all served under a token budget it actually honours. Typical queries cost 50–2,000 tokens against the thousands a naive file read would spend.
Side project — Engineering
A FaceTime MCP for AI agents. It gives an agent a real call as its interface rather than a chat window. Still being built.