Pedro Rafael Castro Sousa
Postgraduate ai researcher and thinkerer who loves learning and figuring out how complex systems work (and occasionally breaking them just to see why). Right now I'm currently learning about zero-cost proxies and MLOps.
Researched reinforcement learning robustness and model safety, designing self-correcting algorithmic frameworks to maintain policy stability under noisy and corrupted environments.
Engineered end-to-end cost forecasting pipelines and built interactive analytics dashboards to predict production costs within 5% of target.
Formulated optimization frameworks to correct corrupted transition data using Bellman error gradients, establishing robust policy performance under high observational uncertainty.
Investigated metric degradation and distribution shifts caused by model interventions, establishing rigorous validation benchmarks and retraining strategies.
Benchmarked deterministic and probabilistic mapping architectures to enhance robotic spatial classification under severe sensor noise.
Faculty of Engineering, University of Porto. Focus on RW-RL frameworks.
Faculty of Sciences, University of Porto. Statistics, ML, and architectures.