An iOS app for critical thinking training. Built solo in six weeks using Anthropic Claude, SwiftUI, and a lot of prompt engineering. The thesis: most AI products optimise for making users think less. This one does the opposite.

A fine grained image classification architecture combining hierarchical attention with gradient based token selection. Lifts top-1 accuracy by 11% while cutting inference latency by 18%. Published at IEEE AAIML 2025. My first real research artefact and the piece of work I'm proudest of methodologically.

An experimental pipeline for fine grained bird identification using vision transformers. Iterated from object detection to a fine grained classification approach to lift accuracy while keeping compute manageable. The lesson I keep returning to: the first architecture that gets you a working result is rarely the right one.

Reinforcement learning applied to autonomous driving simulation. Optimised reward functions to accelerate simulation convergence. Architected real time inference workflows achieving under 150 milliseconds latency. Led the ML strategy for the team.