Built a NEAT (NeuroEvolution of Augmenting Topologies) agent from scratch in C++, evolving both neural-network topology and weights without ML libraries or gradient descent to play a twin-stick space shooter.
Engineered a 41-feature real-time game-state representation and self-growing neural architecture, improving score 52x ($400 \rightarrow 20,800$) and survival time 18x (15s $\rightarrow$ 278s) over 6,000 generations.
Designed a parallelized experiment framework for population-size sweeps and repeated trials across CPU threads, reducing a 4-repeat, 6,000-generation, population-80 sweep to ~12 minutes wall-clock time.
Implemented deterministic seed-based reproducibility across simulation, mutation, crossover, and selection, alongside automated checkpoint/GIF recording, genome visualization, and input-sensitivity analysis across 1,000+ evolved genomes.
