Jonathan’s bespoke project explored reinforcement learning and advanced game AI in Python. He began by developing his own neural network API, giving him the ability to create and train regression models from scratch using any 1D dataset. From there, he progressed into reinforcement learning, designing Q Tables and implementing Deep Q Learning to train agents for tasks such as obstacle avoidance and Snake.

He brought everything together in a spacecraft landing simulation, creating an AI agent capable of teaching itself how to land with increasing precision. Jonathan built the physics, reward shaping, training process and evaluation tools himself, combining mathematical understanding with thoughtful experimentation.

Throughout the project, Jonathan showed significant independence and insight, making thoughtful decisions about design, efficiency, and testing. The result was a university-level AI project that combined mathematical reasoning with real-time game behaviour.

This work became a central focus in Jonathan’s university applications and he has already secured internships at Cisco and Deloitte.